Language understanding and dialogue generation method oriented to power industry

Abstract

The invention provides a power industry-oriented language understanding and dialogue generation method, which comprises the following steps of: retrieving a dialogue process template related to a current problem from a pre-constructed power industry multi-round dialogue knowledge base according to a field label of the problem, and determining a basic framework of a dialogue process and key information required to be acquired by each round of dialogue; after each round of dialogue is finished, the obtained information is synthesized, reasoning is carried out through a pre-constructed electric power knowledge graph, the incidence relation between different professional fields is analyzed, key points and possible solutions of questions are deduced, and candidate answers are generated; and establishing an answer sorting model by applying deep learning, scoring and sorting the candidate answers, and selecting the answer with the highest score as the reply of the current round in combination with the correlation, integrity and specialty of the answers.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a language understanding and dialogue generation method for the power industry. Background Technology

[0002] In real-world business scenarios within the power industry, user questions often involve multiple professional fields, are expressed in diverse ways, and contain a wealth of implicit information. Traditional single-turn question-and-answer models struggle to accurately understand user intent and provide satisfactory responses. To address this issue, a multi-turn dialogue system is needed. This system guides users to clarify their questions through continuous interaction and then comprehensively analyzes the acquired information to provide an accurate answer. However, balancing efficiency and accuracy in multi-turn dialogues presents a significant technical challenge. To improve efficiency, the system needs to minimize the number of dialogue rounds and quickly identify key information; yet, to ensure accuracy, it needs to gather sufficient detailed information through multiple follow-up questions. Furthermore, the system must adjust its question-and-answer strategy in real-time based on user feedback to control the direction of the dialogue and prevent it from going off-topic or reaching dead ends. This requires the system to deeply understand domain knowledge and dynamically optimize dialogue strategies based on that knowledge. Simultaneously, given the highly specialized and complex nature of power industry knowledge, effectively integrating domain knowledge into the optimization process of dialogue strategies is another major challenge. Summary of the Invention

[0003] This invention provides a language understanding and dialogue generation method for the power industry, mainly including:

[0004] By using natural language understanding methods, semantic analysis is performed on the questions raised by users to extract key concepts and professional terms in the power field involved in the questions. Based on the power field ontology knowledge base, the professional field to which the questions belong is inferred, and the field tags of the questions are obtained.

[0005] Based on the domain tags of the problem, retrieve dialogue process templates related to the current problem from the pre-built multi-round dialogue knowledge base of the power industry, and determine the basic framework of the dialogue process and the key information to be obtained in each round of dialogue;

[0006] During the dialogue, the system analyzes the user's input in real time, determines the type of intent, and extracts key information related to the power sector to fill the corresponding slots in the dialogue flow template. Based on the slots in the dialogue flow template, the system obtains the key information needed to fill the slots and gets the next dialogue flow.

[0007] After each round of dialogue, the information obtained is synthesized, and reasoning is performed through a pre-built knowledge graph of electricity to analyze the relationship between different professional fields, infer the key points of the problem and possible solutions, and generate candidate answers.

[0008] A deep learning-based answer ranking model is used to score and rank candidate answers. The answer with the highest score is selected as the response for the current round, taking into account the relevance, completeness, and professionalism of the answers.

[0009] Based on user feedback on the answers, a dialogue strategy model for the power industry is established using reinforcement learning algorithms. The dialogue strategy is adjusted according to the dialogue strategy model to optimize the questioning method and answer generation strategy in the next round. If the user is not satisfied with the answer, follow-up questions are generated according to the dialogue process template to guide the user to provide the necessary details in order to obtain a complete slot filling result.

[0010] If the system still cannot provide a satisfactory answer after multiple rounds of dialogue, it will be transferred to human customer service. The questions, information obtained, and attempted answers from this dialogue will be integrated into a service ticket, which will be followed up and processed by human customer service. The human customer service process will be used as a new sample to optimize the dialogue knowledge base and dialogue strategy model of the power industry dialogue system.

[0011] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0012] This invention discloses a language understanding and dialogue generation method for the power industry. The method extracts key power-related concepts from user questions through semantic analysis, infers the professional domain of the question by combining it with a power ontology knowledge base, and retrieves relevant dialogue process templates from a pre-built multi-turn dialogue knowledge base. During the interaction, the system analyzes user intent in real time, extracts key information to fill template slots, and generates candidate answers based on power knowledge graph reasoning. A deep learning model is used to rank the answers and select the optimal response. If the user is not satisfied, the system adjusts the dialogue strategy based on feedback and optimizes the next round of interaction. If the problem cannot be resolved after multiple rounds of dialogue, the system integrates the dialogue information to generate a work order for manual processing, and uses the processing time to continuously optimize the knowledge base and dialogue model. This invention achieves accuracy, professionalism, and scalability in intelligent question answering in the power field, effectively improving user experience and service efficiency. Attached Figure Description

[0013] Figure 1 This is a flowchart of a language understanding and dialogue generation method for the power industry according to the present invention.

[0014] Figure 2 This is a schematic diagram of a language understanding and dialogue generation method for the power industry according to the present invention.

[0015] Figure 3 This is another schematic diagram of a language understanding and dialogue generation method for the power industry according to the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0017] like Figure 1-3 This embodiment of a language understanding and dialogue generation method for the power industry may specifically include:

[0018] Step S101: Using natural language understanding methods, perform semantic analysis on the questions raised by users, extract key concepts and professional terms in the power field involved in the questions, and infer the professional field to which the questions belong based on the power field ontology knowledge base, thereby obtaining the domain label of the questions.

[0019] The problem text is obtained, and a dictionary- and rule-based named entity recognition method is used to identify the technical terms and key concepts in the problem text, resulting in a standardized problem description. Based on the standardized problem description, the cosine similarity between the problem and multiple preset documents in the power industry domain is calculated using the TF-IDF algorithm, and the domain with the highest similarity is selected as the preliminary classification result. For the preliminary classification result, knowledge nodes of relevant subdomains are selected from the power industry ontology knowledge base to construct a local knowledge graph related to the problem. Concept nodes related to the problem are retrieved from the local knowledge graph, and the association strength between these concept nodes is analyzed using a depth-first search algorithm to determine the core theme of the problem. The preliminary classification result of the problem is fused with the core theme, and a multi-dimensional feature analysis of the problem is performed using the C4.5 decision tree algorithm to calculate the confidence score of each candidate domain label. If there is only one candidate domain label with the highest confidence score, it is determined as the final label of the problem; if the difference in confidence scores among multiple candidate domain labels is within a preset threshold, these multiple candidate domain labels are retained as the domain labels of the problem.

[0020] Specifically, the input question text is segmented, and a dictionary- and rule-based named entity recognition method is used to identify professional terms and key concepts. Corresponding standard expressions are matched from a pre-built power industry dictionary to obtain a standardized question description. Based on the standardized question description, the cosine similarity between the question and multiple pre-set power industry documents is calculated using the TF-IDF algorithm, and the domain with the highest similarity is selected as the preliminary classification result. Based on the preliminary classification result, knowledge nodes of relevant subdomains are selected from the power industry ontology knowledge base to construct a local knowledge graph related to the question. Concept nodes related to the question are retrieved from the constructed local knowledge graph, and the depth-first search algorithm is used to analyze the correlation strength between concepts to determine the core theme of the question. The preliminary classification result of the question is fused with the core theme, and the C4.5 decision tree algorithm is used to perform multi-dimensional feature analysis on the question, calculating a confidence score for each candidate domain label. The label with the highest confidence score is selected as the final label; if multiple label scores are close, multiple possible domain labels are retained, ultimately determining the precise domain label of the question.

[0021] In the text segmentation of the question, the jieba segmentation tool is used for initial segmentation, followed by secondary segmentation optimization using a pre-built professional dictionary in the power industry. The dictionary contains 10,000 commonly used power terms and their synonyms and near-synonyms. Professional terms and key concepts in the text are identified using regular expressions and part-of-speech tagging rules. For example, "distribution network optimization" is identified as a complete professional term. When calculating the similarity between the question and documents in the preset domain using the TF-IDF algorithm, a document set containing 50 power industry professional domains is first established, with each document containing 500 core words from that domain. The TF-IDF value of each word in the question text is calculated to form the feature vector of the question. The same method is used to process the documents in the preset domains to obtain feature vectors for each domain. The cosine similarity between the question vector and the vectors of each domain is calculated, and the highest value is selected as the preliminary classification result. The power industry ontology knowledge base contains 1 million knowledge nodes, covering major domains such as power generation, transmission, distribution, and consumption. Based on the preliminary classification results, 100,000 knowledge nodes from relevant subdomains are selected to construct a local knowledge graph. A depth-first search algorithm is used to analyze the strength of relationships between concepts, with a search depth of 5. The traversal starts at the key concept nodes appearing in the question. The frequency of each node during the traversal is recorded, and the node with the highest frequency is the core topic. When performing multi-dimensional feature analysis using the C4.5 decision tree algorithm, 10 feature dimensions are selected, including word frequency, part-of-speech tagging, and position in the knowledge graph. An information gain ratio threshold of 0.1 is set to generate a decision tree. A confidence score is calculated for each candidate domain label, taking into account factors such as the path length of the decision tree and the purity of leaf nodes. If the difference between the highest and second-highest scores is less than 0.05, these two labels are retained as possible domain labels, ultimately determining the precise domain label for the question.

[0022] Step S102: Based on the domain tags of the problem, retrieve the dialogue process templates related to the current problem from the pre-built multi-turn dialogue knowledge base of the power industry, and determine the basic framework of the dialogue process and the key information to be obtained in each round of dialogue.

[0023] Based on the domain tags of the questions, inverted indexing technology is used to retrieve relevant dialogue process templates from a pre-built multi-turn dialogue knowledge base in the power industry. The top five most similar templates are calculated using cosine similarity as candidate templates. Deep semantic matching is performed on these candidate templates, and the semantic similarity between the questions and keywords in the templates is calculated using the Word2Vec model. The template with the highest semantic similarity is selected as the basic framework for the current dialogue. Key information to be acquired in each round of dialogue is extracted from the selected dialogue process templates, and this key information is labeled using a sequence labeling tool to generate an information acquisition sequence. Based on this information acquisition sequence, a multi-turn dialogue state tracker is constructed to record the information acquired and to be acquired in each round of dialogue. Based on the records of the multi-turn dialogue state tracker, the ID3 decision tree algorithm is used to dynamically adjust the question order and content of each round of dialogue, determining the specific questions and answer strategies for each round. If the acquired information recorded by the multi-turn dialogue state tracker is incomplete, targeted questions are generated to retrieve the missing information. If the acquired information recorded by the multi-turn dialogue state tracker is complete, summary questions are generated to confirm the accuracy of the information. Based on the question sequence, content, and answer strategy, a complete dialogue process plan is developed to ensure the coherence of the dialogue and the completeness of information acquisition.

[0024] Specifically, based on the domain tags of the questions, an inverted index technique is used to quickly retrieve relevant dialogue process templates from a pre-built multi-turn dialogue knowledge base in the power industry. The top five most similar templates are then selected as candidates using cosine similarity calculation. Deep semantic matching is performed on the candidate templates, and the semantic similarity between the questions and keywords in the templates is calculated using the Word2Vec model. The template with the highest semantic similarity is selected as the basic framework for the current dialogue. Key information to be acquired in each round of dialogue is extracted from the selected dialogue process templates, and sequence labeling tools are used to annotate the key information, generating an information acquisition sequence. Based on the information acquisition sequence, a multi-turn dialogue state tracker is constructed to record the information acquired and to be acquired in each round of dialogue, providing a basis for subsequent dialogue process adjustments. Based on the records of the multi-turn dialogue state tracker, the ID3 decision tree algorithm is used to dynamically adjust the question order and content of each round of dialogue, ensuring the coherence of the dialogue and the completeness of information acquisition. Simultaneously, the specific questions and answer strategies for each round of dialogue are determined based on the specific content of the questions, forming a complete dialogue process plan. The power industry multi-turn dialogue knowledge base pre-stores 10,000 dialogue process templates, each containing an average of 5 rounds of dialogue. Using an inverted index technique with the domain tags of the question as search keywords, 50 relevant templates are retrieved from the knowledge base within 0.1 seconds. Cosine similarity calculation is used to select the top 5 templates with a similarity greater than 0.8 as candidates. A pre-trained Word2Vec model containing 1 million 300-dimensional word vectors is used to calculate the semantic similarity between the question and the keywords in the candidate templates. The template with the highest semantic similarity is selected, with a similarity threshold set to 0.9. Key information is extracted from the selected templates, and a sequence labeling tool based on a bidirectional long short-term memory network is used, achieving 95% accuracy, to generate an information retrieval sequence containing 10 key information points. A multi-turn dialogue state tracker is constructed, using a hash table to store retrieved and pending information, with an update time complexity of O. The ID3 decision tree algorithm is used, with information gain as the splitting criterion, a tree depth limit of 5 layers, and a minimum number of samples per leaf node of 10, to dynamically generate the decision tree. Decision tree nodes contain attributes such as dialogue turn, retrieved information, and pending information, selecting the optimal content for the next round of dialogue based on the current state. The final generated dialogue flow plan includes 6 rounds of dialogue, with an average of 3 questions per round, covering 95% of the key information points.

[0025] Step S103: During the dialogue, the user's input is analyzed in real time to determine the type of intent and extract key information related to the power sector, which is then used to fill the corresponding slots in the dialogue flow template. Based on the slots in the dialogue flow template, the necessary key information for filling is obtained to obtain the next dialogue flow.

[0026] User input information is received using streaming processing technology. This input information is then processed by natural language processing tools for word segmentation and part-of-speech tagging. Based on the segmentation and tagging results, a support vector machine is used to classify the user intent, determining the intent type. Simultaneously, a conditional random field algorithm is used to extract key information related to the power sector from the user input. For this key information, corresponding slots are located in a pre-established dialogue flow template. An edit distance algorithm is used for fuzzy matching to fill the corresponding slots with the key information. A recurrent neural network is used to prioritize filled and unfilled slots, and combined with the current dialogue context, the next key information content to be acquired is determined. Based on the next key information content, a question is selected from a pre-set question template library and adjusted according to the current dialogue context to obtain a personalized question.

[0027] Specifically, the system employs streaming technology to receive and process user input in real time. Natural language processing tools are used for word segmentation and part-of-speech tagging. Support vector machines are used for intent classification to determine the user's intent type. Simultaneously, a conditional random field algorithm is used to extract key information related to the power industry from the user input. Based on the extracted key information, corresponding slots are located in the dialogue flow template. An edit distance algorithm is used for fuzzy matching to fill the corresponding slots with the key information, and unfilled slots are marked. A recurrent neural network is used to prioritize filled and unfilled slots, and combined with the current dialogue context, the key information to be obtained next is determined. Based on the determined key information to be obtained next, a suitable question is selected from a pre-set question template library, and the question is personalized according to the current dialogue context. Based on the slot filling status, priority ranking results, and personalized question, the most suitable next dialogue content is selected from the dialogue flow template, and the optimal dialogue flow path is generated using the Viterbi algorithm. In the power industry dialogue system, Kafka streaming technology is used to receive user input in real time, processing 1000 messages per second. The jieba word segmentation tool is used to segment the input, achieving a part-of-speech tagging accuracy of 95%. Intent classification was performed using a linear kernel support vector machine (SVM) with 10,000 labeled training data points, achieving a classification accuracy of 90%. Simultaneously, a conditional random field (CRF) algorithm was used to extract key information, achieving an F1 score of 0.85. The dialogue flow template contained 50 predefined slots, and fuzzy matching was performed using the Levenshtein distance algorithm with a threshold of 0.8. A bidirectional long short-term memory (LSTM) network was used to prioritize the slots; the network contained two hidden layers, each with 128 neurons, and underwent 50 training epochs. The question template library contained 1000 preset questions, and the most similar question was selected using a cosine similarity algorithm with a similarity threshold of 0.7. Personalized adjustments were performed using a rule base containing 200 adjustment rules. Finally, the Viterbi algorithm was used to generate the optimal dialogue flow path, with a 100x100 state transition matrix and a 100x50 observation probability matrix.

[0028] Based on user input, an intent recognition model is used to determine the type of dialogue intent. Key information is then extracted based on the intent type and filled into the corresponding dialogue flow template slots. If the slots are not fully filled, follow-up questions are generated based on the dialogue flow template to guide the user to provide the necessary details in order to obtain a complete slot filling result and determine the next step of the dialogue flow.

[0029] The user input is segmented and labeled using natural language processing (NLP) techniques, and the user intent type is determined by a pre-trained support vector machine (SVM) dialogue intent recognizer. Based on the user intent type, key information related to the power sector is extracted from the user input using a conditional random field (CRF) algorithm. After obtaining the user intent type and the key information, the corresponding slots are located in the dialogue flow template, and the key information is filled into the corresponding slots using the Levenshtein distance algorithm. If the slot filling ratio is lower than a preset threshold, the filled and unfilled slots are prioritized using a long short-term memory (LSTM) network, and follow-up questions are selected from a preset question template library. Based on the slot filling status, priority ranking results, and follow-up questions, the optimal dialogue path is calculated using the Viterbi algorithm to obtain the next dialogue flow step.

[0030] Specifically, natural language processing (NLP) techniques are used to segment and tag user input. A pre-trained support vector machine (SVM) dialogue intent recognizer determines the user's intent type, and a conditional random field (CRF) algorithm extracts key information related to the power sector from the user input. Based on the identified intent type and extracted key information, corresponding slots are located in the dialogue flow template. The Levenshtein distance algorithm is used to fill the corresponding slots with the key information, and unfilled slots are marked. A filling ratio threshold is set to determine whether the slots are fully filled; if the filling ratio is below the threshold, they are marked as incomplete. A long short-term memory (LSTM) network is used to prioritize filled and unfilled slots. Combined with the current dialogue context, appropriate follow-up questions are selected from a pre-set question template library, and targeted guiding questions are generated through template filling. The generated guiding questions are added to the dialogue state as the output of the next round of dialogue, and the dialogue history is updated. Based on the slot filling state, priority ranking results, and generated guiding questions, the Viterbi algorithm is used to calculate the optimal dialogue path. If the slots are completely filled, the next dialogue step is determined directly; if incomplete, the generated guiding question is used as the next dialogue content, and the result is stored in the dialogue state manager. In the intelligent customer service system for the power industry, user input is processed by the jieba word segmentation tool, identifying an average of 15 words with a part-of-speech tagging accuracy of 95%. The pre-trained support vector machine intent recognizer, based on 10,000 labeled data points, uses the RBF kernel function with a C value of 1.0 and a gamma value of 0.1, achieving a classification accuracy of 92% for 20 common intents. The conditional random field algorithm uses a feature template containing 10 features including words, parts of speech, and location, achieving an F1 score of 0.88 for extracting key information. The dialogue flow template contains 50 predefined slots, which are filled using the Levenshtein distance algorithm with a distance threshold of 3. The filling ratio threshold is set to 70%; anything below this value is marked as incomplete. The Long Short-Term Memory (LSTM) network consists of two hidden layers, each with 128 neurons, and a dropout rate of 0.5. After 50 training epochs, it achieves 85% accuracy in slot priority ranking. The question template library contains 1000 preset questions. Semantic similarity is calculated using Word2Vec, and the top three questions with a similarity greater than 0.8 are selected for template filling. The dialogue state manager uses Redis caching to store the history of the most recent 5 rounds of dialogue.

[0031] The Viterbi algorithm uses a 10x10 state transition matrix and a 10x50 emission probability matrix to calculate the optimal path, with an average time of 20 milliseconds. The average response time of the entire process is controlled within 300 milliseconds, meeting the requirements for real-time interaction. The system can handle 95% of common power service queries, covering major scenarios such as fault reporting, electricity bill inquiry, and service processing.

[0032] Step S104: After each round of dialogue, the information obtained is synthesized, and reasoning is performed through a pre-built knowledge graph of electricity to analyze the relationship between different professional fields, infer the key points of the problem and possible solutions, and generate candidate answers.

[0033] Natural language processing (NLP) techniques are used to segment and identify entities from the acquired information, resulting in a set of identified entities. Based on this set, a depth-first search algorithm is used to search for paths between related entities in a pre-constructed power knowledge graph to determine the strength of the association between entities. If the association strength is greater than a preset threshold, the PageRank algorithm is used to calculate the importance score of each entity node, and the top three entities with the highest scores are selected as the key points of the problem. For the key points of the problem, nodes and edges directly connected to the key points are extracted from the power knowledge graph to construct a subgraph as the key area for reasoning. Based on the subgraph, the Rete algorithm is used for rule-based reasoning, extracting relevant facts and rules from the power knowledge graph, and generating a solution through forward linking reasoning.

[0034] Specifically, by setting a dialogue state flag, when the user completes input and the system response ends, the dialogue turn flag is set to the "ended" state, triggering subsequent processing. Natural language processing (NLP) techniques are used to segment and identify entities from the acquired information. Entities are linked using a pre-constructed power knowledge graph, mapping the identified entities to nodes in the knowledge graph. A depth-first search algorithm is used to search for paths between related entities in the knowledge graph. The association strength between entities is calculated based on path length and edge weight, revealing the relationships between different professional fields. Based on the association analysis results, the PageRank algorithm is used to calculate the importance score of each entity node, selecting the top three entities with the highest scores as the key points of the problem. Based on the identified key points, nodes and edges directly connected to these key points are extracted from the knowledge graph, constructing a subgraph as the focus area for reasoning. Based on the constructed subgraph, the Rete algorithm is used for rule-based reasoning, extracting relevant facts and rules from the knowledge graph and generating possible solutions through forward linking reasoning. Combining multi-turn dialogue history information, the generated solutions are ranked using cosine similarity calculation. A weighted average method is used to calculate the final score of the candidate answers, and the top five with the highest scores are selected as candidate answers. In the intelligent power customer service system, the dialogue status manager uses Redis caching, setting the key-value pair "dialogue_status" to 1 to indicate that the dialogue is in progress and 0 to indicate that the turn has ended. After each user input is processed, the status is set to 0, triggering the subsequent process. The jieba word segmentation tool is used to process the acquired information, identifying an average of 20 entities. Entity links are established using a pre-constructed power knowledge graph containing 1 million nodes and 5 million edges, achieving a linking accuracy of 95%. The depth-first search algorithm is set to a maximum depth of 5, a search timeout of 2 seconds, and an average execution time of 150 milliseconds. The strength of the association between entities is calculated by multiplying the inverse of the path length by the edge weight, with the weight ranging from 0 to 1.

[0035] The PageRank algorithm iterates 10 times with a damping coefficient of 0.85 and a convergence threshold of 0.0001, selecting the three entities with the highest scores as keypoints. Nodes and edges directly connected to the keypoints are extracted from the knowledge graph, constructing a subgraph with an average of 50 nodes and 200 edges. The Rete algorithm processes 1000 predefined rules with an inference depth of 3, generating 10 possible solutions. Combining the most recent 5 rounds of dialogue history, TF-IDF is used to calculate word frequency features, with a cosine similarity threshold of 0.7, and the solutions are ranked. An exponentially decaying weighted average is used, with the most recent round having a weight of 0.5, decreasing sequentially, to calculate the final score of the candidate answers, selecting the five with the highest scores as output.

[0036] Step S105: Use deep learning to build an answer ranking model, score and rank the candidate answers, and select the answer with the highest score as the response for the current round, based on the relevance, completeness and professionalism of the answers.

[0037] A pre-trained word vector model, incorporating Word2Vec word embedding technology, is used to vectorize the candidate answer text. Semantic features of the candidate answer text are obtained, and feature extraction is performed using a convolutional neural network based on these features, resulting in a low-dimensional vector representation of the candidate answer. This low-dimensional vector representation is concatenated with the question vector and dialogue history vector to obtain a unified input vector. A multilayer perceptron network (MLB) is constructed based on this unified input vector. The MLB consists of three layers, each containing a predetermined number of neurons. Nonlinear transformations are performed using fully connected layers and the ReLU activation function of the MLB to output relevance, completeness, and professionalism scores for the candidate answer. A self-attention mechanism is used to weight key information in the candidate answer. Professional terms are retrieved from a pre-established dictionary and terminology table for the power industry. The professionalism score of the candidate answer is calculated based on the weighted key information and the professional terms.

[0038] Specifically, Word2Vec word embedding technology is used to vectorize candidate answers. A pre-trained word vector model is used to extract semantic features from the answer text. A convolutional neural network is then used to further extract and combine these features to obtain a low-dimensional vector representation of the answer. The answer vector, question vector, and dialogue history vector are concatenated into a unified input vector, which serves as the input to a multilayer perceptron network. A three-layer perceptron network is constructed as the basic structure of the answer ranking model, with each layer containing 256, 128, and 64 neurons respectively. Nonlinear transformations are performed using fully connected layers and the ReLU activation function to output the relevance, completeness, and professionalism scores of the answers. The multilayer perceptron network is trained on a dataset of 1 million labeled historical question-and-answer entries, employing a cross-entropy loss function and the Adam optimizer, with a learning rate of 0.001, for 50 training epochs. A self-attention mechanism is used to weight key information in the answers. Combined with a dictionary of 100,000 terms in the power industry and a glossary of professional terms, the professionalism score of the answer is calculated. Simultaneously, relevance and completeness scores are calculated based on the cosine similarity between the question and answer and information coverage. The relevance, completeness, and professionalism scores are weighted and summed to obtain a comprehensive score for the answer. A quick sorting algorithm is used to rank the candidate answers, and the answer with the highest score is selected as the response for the current round. The selected answer and its score are stored in the dialogue history. User feedback data is collected regularly, and the model parameters are updated weekly using incremental learning methods to adapt to new question-answering patterns and changes in domain knowledge. In the intelligent customer service system for the power industry, a pre-trained Word2Vec model containing 1 million 300-dimensional word vectors is used to vectorize the candidate answers. A 1D convolutional neural network with three convolutional layers, each with 64, 128, and 256 convolutional kernels of sizes 3, 4, and 5 respectively, is used to obtain a 128-dimensional vector representation of the answer. The answer vector, question vector, and the vectors from the last five rounds of dialogue history are concatenated into a unified 512-dimensional input vector. A three-layer perceptron network is constructed, with 256, 128, and 64 neurons per layer, using the ReLU activation function. The output layer uses the Softmax function to obtain three scores. The training dataset contains 1 million labeled question-answer pairs with a batch size of 64. The Adam optimizer is used with an initial learning rate of 0.001, decaying by 10% every 10 epochs for a total of 50 epochs. The validation set accuracy reaches 92%. The self-attention mechanism uses 8 attention heads with a hidden layer dimension of 64. The power industry dictionary contains 100,000 terms, and the TF-IDF algorithm is used to calculate the weights of specialized terms. Relevance scores are calculated using cosine similarity, completeness scores are based on keyword coverage, and professionalism scores are calculated based on the proportion of specialized terms. The weights for the three scores are 0.4, 0.3, and 0.3, respectively. A quicksort algorithm is used to sort candidate answers, with an average sorting time of 10 milliseconds.We collect 5,000 user feedback data points each week and use stochastic gradient descent for incremental learning to update the model parameters.

[0039] By acquiring the knowledge domain and question type attributes of candidate answers, a semantic similarity model is used to determine the relevance of the answer to the question. The completeness of the answer is determined based on the coverage of key information, and the professionalism score of the answer is obtained by combining the professional vocabulary ratio index. The answer with the highest professionalism score is obtained by comprehensive ranking.

[0040] A pre-trained BERT text classifier is used to classify candidate answers by knowledge domain and label them by question type, resulting in classification results. Based on the classification results, the semantic similarity score between the candidate answer and the question is calculated using the Word2Vec word vector model to obtain the relevance score of the candidate answer. Dependency parsing is used to extract the core entities and relationships in the question, and an information coverage checklist is constructed. The completeness score of the candidate answer is calculated using the checklist. A professional dictionary in the power industry is used to identify professional terms in the candidate answers, and the frequency and proportion of professional terms in the candidate answers are statistically analyzed to calculate the professionalism score of the candidate answers. The candidate answers are sorted using a heap sort algorithm. If the comprehensive score of the candidate answers is higher than a preset threshold, the candidate answer with the highest comprehensive score is selected as the final output result.

[0041] Specifically, a pre-trained BERT text classifier is used to classify candidate answers by knowledge domain and label them by question type. Keywords for both answers and questions are extracted using the term frequency-inverse document frequency (TF-IDF) algorithm, and the semantic similarity score between answers and questions is calculated using the Word2Vec word vector model. Based on the calculated semantic similarity score, a threshold of 0.7 is set to filter out candidate answers with high relevance for the next step of completeness assessment. Dependency parsing is used to extract core entities and relationships from the questions, constructing an information coverage checklist. Based on the checklist, the number of key information items contained in the answers is matched using regular expressions to calculate the information coverage rate and obtain the completeness score of the answer. A professional dictionary in the power industry is used to identify professional vocabulary in the answers, and the frequency and proportion of professional vocabulary in the answers are statistically analyzed. The importance weight of professional vocabulary is calculated using the TF-IDF algorithm to calculate the professionalism score of the answer. The relevance score, completeness score, and professionalism score are weighted and summed. In the comprehensive score calculation, the professionalism score is given a weight of 0.5 to ensure that highly professional answers are more likely to be selected. The heap sort algorithm is used to sort the candidate answers, and the answer with the highest comprehensive score is selected as the final output. In the intelligent customer service system for the power industry, a pre-trained BERT-based model is used as the text classifier. This model contains 120 million parameters and has been fine-tuned on 10 million pieces of text in the power industry field, achieving classification across 20 knowledge domains and 50 question types with an accuracy of 95%. The TF-IDF algorithm is used to extract the top-10 keywords, and the Word2Vec model, using 300-dimensional word vectors, is trained on a 100 million-word power industry corpus. Semantic similarity is calculated using cosine similarity with a threshold of 0.7, selecting an average of 80% relevant answers. Dependency parsing uses the Stanford CoreNLP tool to extract 5-8 core entities and relations to construct a checklist. Regular expression matching uses the Python re library with a coverage threshold of 70%. The power industry professional dictionary contains 100,000 terms, and TF-IDF weight calculation is implemented using the scikit-learn library. The weights for relevance, completeness, and professionalism scores are 0.3, 0.2, and 0.5, respectively, and weighted summation is performed using the NumPy library. Heap sort, using the Python heapq module, sorts an average of 100 candidate answers in less than 10 milliseconds.

[0042] Step S106: Based on user feedback on the answers, a dialogue strategy model for the power industry is established using a reinforcement learning algorithm. The dialogue strategy is then adjusted according to the model to optimize the questioning method and answer generation strategy for the next round. If the user is not satisfied with the answer, follow-up questions are generated based on the dialogue flow template to guide the user to provide the necessary details to obtain a complete slot filling result.

[0043] A dialogue strategy model is constructed using the DeepQ-Network algorithm. This model includes a state space and an action space specific to the power industry. The state space includes user type, question category, and dialogue history, while the action space includes question type, answer detail level, and the use of technical terms. The dialogue strategy model is trained using a state-action-reward approach, where the reward is obtained from user feedback. The optimal action is selected based on the Q-value output by the dialogue strategy model. This optimal action determines the specific way to ask questions in the next round, including the wording, tone, and complexity of the question, and also optimizes the parameters for answer generation, including the level of detail and the frequency of technical terminology. If the user is not satisfied with the answer, a sequence-to-sequence model based on an attention mechanism is used to generate follow-up questions. These follow-up questions are selected and generated from a predefined follow-up question template library based on the current dialogue state and unfilled slots. Named entity recognition and dependency parsing techniques are used to analyze the user's responses to the follow-up questions. These techniques extract key information and fill it into the corresponding slots, updating the dialogue state. The first two steps are repeated until a complete slot filling result is obtained or a preset maximum number of rounds is reached.

[0044] Specifically, a Deep Q-Network algorithm is used to construct a dialogue strategy model. A state space specific to the power industry is defined, including user type, question category, and dialogue history. The action space includes question type, answer detail level, and use of technical terms. User feedback is used as a reward signal. The model is trained through a state-action-reward approach to optimize the dialogue strategy. Based on the Q-value output by the dialogue strategy model, the optimal action is selected to determine the specific way to ask questions in the next round, including the wording, tone, and complexity of the question. Simultaneously, the parameters for answer generation are optimized, such as adjusting the level of detail and the frequency of technical terminology. If the user is not satisfied with the answer, a sequence-to-sequence model based on an attention mechanism is used. Based on the current dialogue state and unfilled slots, the most suitable follow-up question is selected and generated from a predefined follow-up question template library. Named entity recognition and dependency parsing techniques are used to analyze the user's response to the follow-up question, extract key information, and fill it into the corresponding slots. The dialogue state is updated, and the first two steps are repeated until a complete slot filling result is obtained or the preset maximum number of rounds is reached.

[0045] In the intelligent customer service system for the power industry, the Deep Q-Network algorithm is used to construct the dialogue strategy model. The state space dimension is set to 100, including 5 user types, 20 question categories, and 75 dialogue history features. The action space dimension is 50, covering 10 question types, 5 levels of answer detail, and 35 options for using technical terms. The reward function is based on user feedback, with satisfaction scores ranging from 0 to 10, converted into reward values ​​from -1 to 1. The model uses a three-layer fully connected network with 256 and 128 neurons in the hidden layers, respectively. The learning rate is set to 0.001, the discount factor is 0.95, and the ε-greedy strategy has an initial ε value of 0.9, decaying at a rate of 0.995. The target network is updated every 1000 rounds of dialogue. The sequence-to-sequence model uses a bidirectional LSTM structure, with two layers each for the encoder and decoder, 512 hidden units, and Bahdanau attention as the attention mechanism. The follow-up question template library contains 1000 templates, covering 50 common power business scenarios. Named entity recognition uses a BiLSTM-CRF model, trained on 100,000 labeled data points, achieving an F1 score of 0.92. Dependency parsing employs a transition-based neural network parser, achieving a UAS accuracy of 92%. The entire system runs on a server with a 32-core CPU and 128GB of memory, with an average response time controlled within 500 milliseconds, processing 200,000 dialogues per day, and improving user satisfaction from an initial 75% to 90%.

[0046] Step S107: If the system still cannot provide a satisfactory answer to the user after multiple rounds of dialogue, it will be transferred to human customer service for processing. The questions, information obtained, and attempted answers from this dialogue will be integrated into a service ticket, which will be followed up and processed by human customer service. The human customer service process will be used as a new sample to optimize the dialogue knowledge base and dialogue strategy model of the power industry dialogue system.

[0047] The system receives service request information carrying a dialogue round count and user satisfaction evaluation indicators, generated when a user initiates an electricity business inquiry. Based on the service request information, it determines whether the dialogue rounds exceed a preset threshold or whether user satisfaction is continuously below a set value. If either threshold is exceeded or the value is below the set value, a human customer service transfer process is triggered. Named entity recognition and dependency parsing techniques are used to extract information from the multi-turn dialogue content, obtaining a user problem description and key information points. A text summarization algorithm is used to process the user problem description and key information points, generating a structured service ticket. Based on the electricity business type and customer service expertise information in the service ticket, the ticket is automatically assigned to the corresponding human customer service workstation through the ticket management system. Real-time speech-to-text technology is used to record the human customer service process, obtaining processing flow data. Long Short-Term Memory (LSTM) networks are used to analyze the processing flow data, extracting key decision points and processing strategies. The knowledge base of the electricity industry dialogue system is updated based on the key decision points and processing strategies, optimizing the dialogue strategy model.

[0048] Specifically, a dialogue round counter and a user satisfaction evaluation mechanism based on power industry customer service standards are set up, including satisfaction evaluation indicators such as service attitude, problem-solving level, and professionalism. When the number of dialogue rounds exceeds a preset threshold or user satisfaction continuously falls below a set value, a human customer service transfer process is triggered. Named entity recognition and dependency parsing technologies are used to extract information from multi-turn dialogue content, and text summarization algorithms are used to generate problem descriptions, integrating them into structured service tickets, including user problem descriptions, key information points, and answers tried by the system. The service ticket management system automatically assigns service tickets based on power business type and customer service expertise, and pushes the service tickets to the customer service workbench. At the same time, real-time speech-to-text technology records the human customer service process. The human customer service process is analyzed through long short-term memory network analysis to extract key decision points and processing strategies, automatically updating the knowledge base of the power industry dialogue system. The extracted processing strategies are used to optimize the dialogue strategy model, improving the system's ability to handle complex power business problems. In the intelligent customer service system for the power industry, the dialogue round counter is set to a threshold of 10 rounds. User satisfaction is evaluated using a 5-point scale, including three dimensions: service attitude, problem-solving level, and professionalism, with weights of 0.3, 0.5, and 0.2 respectively. A human agent transfer is triggered when the score is below 3.5 for three consecutive rounds. Named entity recognition uses a BiLSTM-CRF model, trained on 100,000 labeled data points in the power industry, achieving an F1 score of 0.94. Dependency parsing uses a graph neural network-based parser, achieving a User-Agent Search (UAS) accuracy of 93%. Text summarization uses a Transformer-based BART model, fine-tuned on 50,000 power industry customer service dialogues, achieving a ROUGE-L score of 0.72. The work order management system uses a rule engine to match 20 power industry business types and 50 customer service specialty tags, with an average allocation time of less than 100ms. Speech-to-text conversion uses a Transformer-based end-to-end ASR model, trained on 10,000 hours of power industry customer service call data, with a word error rate of less than 5%. An LSTM network was used to analyze customer service processes. It consisted of a 3-layer bidirectional LSTM with 256 hidden units, trained for 50 epochs on 100,000 manual processing records, extracting an average of 8 key decision points. The knowledge base was updated using an incremental learning method, summarizing new samples daily and updating in batches weekly. The dialogue strategy model was fine-tuned using the PPO algorithm on the updated knowledge base, iterating 10,000 steps per iteration.

[0049] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A language understanding and dialogue generation method for the power industry, characterized in that, The method includes: By using natural language understanding methods, semantic analysis is performed on the questions raised by users to extract key concepts and professional terms in the power field involved in the questions. Based on the power field ontology knowledge base, the professional field to which the questions belong is inferred, and the field tags of the questions are obtained. Based on the domain tags of the problem, retrieve dialogue process templates related to the current problem from the pre-built multi-round dialogue knowledge base of the power industry, and determine the basic framework of the dialogue process and the key information to be obtained in each round of dialogue; During the dialogue, the system analyzes the user's input in real time, determines the type of intent, and extracts key information related to the power sector to fill the corresponding slots in the dialogue flow template. Based on the slots in the dialogue flow template, the system obtains the key information needed to fill the slots and gets the next dialogue flow. After each round of dialogue, the information obtained is synthesized, and reasoning is performed through a pre-built knowledge graph of electricity to analyze the relationship between different professional fields, infer the key points of the problem and possible solutions, and generate candidate answers. A deep learning-based answer ranking model is used to score and rank candidate answers. The answer with the highest score is selected as the response for the current round, taking into account the relevance, completeness, and professionalism of the answers. Based on user feedback on the answers, a dialogue strategy model for the power industry is established using reinforcement learning algorithms. The dialogue strategy is adjusted according to the dialogue strategy model to optimize the questioning method and answer generation strategy in the next round. If the user is not satisfied with the answer, follow-up questions are generated according to the dialogue process template to guide the user to provide the necessary details in order to obtain a complete slot filling result. If the system still cannot provide a satisfactory answer after multiple rounds of dialogue, it will be transferred to human customer service. The questions, information obtained, and attempted answers from this dialogue will be integrated into a service ticket, which will be followed up and processed by human customer service. The human customer service process will be used as a new sample to optimize the dialogue knowledge base and dialogue strategy model of the power industry dialogue system.

2. The method according to claim 1, characterized in that, The method employs natural language understanding to perform semantic analysis on user-submitted questions, extracting key concepts and technical terms related to the power industry. Based on a power industry ontology knowledge base, it infers the professional field to which the question belongs, obtaining the question's domain label, including: The problem text is obtained, and the professional terms and key concepts in the problem text are identified using a dictionary- and rule-based named entity recognition method to obtain a standardized problem description. Based on the standardized problem description, the cosine similarity between the problem and multiple preset documents in the power industry field is calculated using the TF-IDF algorithm, and the field with the highest similarity is selected as the preliminary classification result of the problem. Based on the preliminary classification results, knowledge nodes of relevant subdomains are selected from the power domain ontology knowledge base to construct a local knowledge graph related to the problem. Concept nodes related to the problem are retrieved from the local knowledge graph, and the strength of the association between the concept nodes is analyzed using a depth-first search algorithm to determine the core theme of the problem. The preliminary classification results of the problem are integrated with the core theme, and the problem is subjected to multi-dimensional feature analysis using the C4.5 decision tree algorithm to calculate the confidence score of each candidate domain label. If there is only one candidate domain label with the highest confidence score, then that label is determined as the final label for the problem; If the confidence scores of multiple candidate domain labels are within a preset threshold, then the multiple candidate domain labels are retained as the domain labels of the problem.

3. The method according to claim 1, characterized in that, Based on the domain tags of the question, a dialogue process template related to the current question is retrieved from a pre-built multi-turn dialogue knowledge base in the power industry. This determines the basic framework of the dialogue process and the key information to be obtained in each round of dialogue, including: Based on the domain tags of the problem, inverted indexing technology is used to retrieve relevant dialogue process templates from a pre-built multi-turn dialogue knowledge base in the power industry. The five most similar templates are then calculated using cosine similarity as candidate templates. Deep semantic matching is performed on the candidate templates. The semantic similarity between the question and the keywords in the template is calculated using the Word2Vec model. The template with the highest semantic similarity is selected as the basic framework of the current dialogue. Extract the key information to be obtained in each round of dialogue from the selected dialogue flow template, and annotate the key information using a sequence labeling tool to generate an information acquisition sequence; Based on the information acquisition sequence, a multi-turn dialogue state tracker is constructed to record the information acquired and to be acquired in each round of dialogue. Based on the records of the multi-turn dialogue state tracker, the ID3 decision tree algorithm is used to dynamically adjust the question order and content of each round of dialogue, and to determine the specific questions and answer strategies for each round of dialogue. If the information recorded by the multi-turn dialogue state tracker is incomplete, a targeted question is generated to retrieve the missing information. If the acquired information recorded by the multi-turn dialogue state tracker is complete, a summary question is generated to confirm the accuracy of the information. Based on the question sequence, content, and answering strategies, a complete dialogue process plan is developed to ensure the coherence of the dialogue and the completeness of information acquisition.

4. The method according to claim 1, characterized in that, During the dialogue, the system analyzes user input in real time, determines the type of intent, and extracts key information related to the power sector. This information is then used to fill the corresponding slots in the dialogue flow template. Based on the slots in the template, the system obtains the necessary key information to fill them, leading to the next step in the dialogue flow, including: User input information is received using streaming processing technology, and the user input information is processed by natural language processing tools for word segmentation and part-of-speech tagging. Based on the results of word segmentation and part-of-speech tagging, the user intent is classified using a support vector machine to obtain the user intent type. At the same time, key information in the power sector is extracted from the user input information using a conditional random field algorithm. For the key information in the power sector, the corresponding slots are found in the pre-established dialogue process template, and the edit distance algorithm is used for fuzzy matching to fill the key information in the power sector into the corresponding slots; A recurrent neural network is used to prioritize filled and unfilled slots, and combined with the current dialogue context information, the key information to be obtained next is determined. Based on the key information obtained in the next step, a question is selected from the preset question template library, and the question is adjusted according to the current dialogue context information to obtain a personalized question; It also includes: based on user input, using an intent recognition model to determine the type of dialogue intent, obtaining key information based on the intent type, filling the extracted key information into the corresponding dialogue flow template slots, and generating corresponding follow-up questions based on the dialogue flow template to guide the user to provide the necessary details in order to obtain a complete slot filling result and determine the next step of the dialogue flow.

5. The method according to claim 4, characterized in that, The process involves using an intent recognition model based on user input to determine the dialogue intent type, extracting key information based on the intent type, and filling the extracted key information into the corresponding dialogue flow template slots. If the slots are not fully filled, follow-up questions are generated based on the dialogue flow template to guide the user to provide the necessary details to obtain a complete slot filling result and determine the next step of the dialogue flow, including: Natural language processing techniques are used to segment and tag part-of-speech tags of user input, and a pre-trained support vector machine dialogue intent recognizer is used to determine the type of user intent. Based on the user intent type, key information related to the power sector is extracted from the user input using a conditional random field algorithm. Obtain the user intent type and the key information, find the corresponding slot in the dialogue flow template, and use the Levenshtein distance algorithm to fill the key information into the corresponding slot; If the slot filling ratio is lower than the preset threshold, the long short-term memory network is used to prioritize the filled and unfilled slots, and follow-up questions are selected from the preset question template library. Based on the slot filling status, priority sorting results, and follow-up questions, the Viterbi algorithm is used to calculate the optimal dialogue path and obtain the next dialogue flow.

6. The method according to claim 1, characterized in that, After each round of dialogue, the system synthesizes the acquired information, uses a pre-constructed knowledge graph of the power industry to reason, analyzes the relationships between different professional fields, infers the key points of the problem and possible solutions, and generates candidate answers, including: Natural language processing techniques are used to segment and recognize entities from the acquired information to obtain a set of recognized entities. Based on the identified entity set, a depth-first search algorithm is used to search for paths between related entities in a pre-constructed power knowledge graph to determine the strength of the association between entities; If the association strength is greater than a preset threshold, the PageRank algorithm is used to calculate the importance score of each entity node, and the top three entities with the highest scores are selected as the key points of the problem. To address the key points of the problem, nodes and edges directly connected to the key points are extracted from the power knowledge graph, and a subgraph is constructed as the key area for reasoning. Based on the subgraph, the Rete algorithm is used for rule-based reasoning to extract relevant facts and rules from the power knowledge graph, and a solution is generated through forward linking reasoning.

7. The method according to claim 1, characterized in that, The process involves using deep learning to build an answer ranking model, scoring and ranking candidate answers, and selecting the highest-scoring answer as the response for the current round, based on the answers' relevance, completeness, and professionalism. The candidate answer text is vectorized using a pre-trained word vector model, which includes Word2Vec word embedding technology; Obtain the semantic features of the candidate answer text, and perform feature extraction using a convolutional neural network based on the semantic features; Obtain a low-dimensional vector representation of the candidate answer; The low-dimensional vector representation of the candidate answer is concatenated with the question vector and the dialogue history vector to obtain a unified input vector; A multilayer perceptron network is constructed based on the unified input vector. The multilayer perceptron network includes a three-layer structure, with each layer containing a predetermined number of neurons. The nonlinear transformation is performed through the fully connected layer and ReLU activation function of the multilayer perceptron network to output the relevance, completeness and professionalism scores of the candidate answers; The key information in the candidate answers is weighted using a self-attention mechanism; Obtain specialized terms from a pre-established dictionary and glossary of terms in the power industry; The professionalism score of the candidate answers is calculated based on the weighted key information and the professional terminology used. It also includes: obtaining the knowledge domain and question type attributes of candidate answers, using a semantic similarity model to determine the relevance of the answer to the question, determining the completeness of the answer based on the coverage of key information, and combining the professional vocabulary ratio index to obtain the professional score of the answer, and comprehensively ranking to obtain the answer with the highest professional score.

8. The method according to claim 7, characterized in that, The process involves acquiring the knowledge domain and question type attributes of candidate answers, using a semantic similarity model to determine the relevance of the answers to the questions, determining the completeness of the answers based on the coverage of key information, and combining this with a professional vocabulary ratio to obtain a professionalism score. The answers with the highest professionalism scores are then ranked comprehensively. This includes: A pre-trained BERT text classifier is used to classify candidate answers by knowledge domain and label them by question type, thus obtaining the classification results of the candidate answers; Based on the classification results, the semantic similarity score between the candidate answer and the question is calculated using the Word2Vec word vector model to obtain the relevance score of the candidate answer; Dependency parsing is used to extract the core entities and relationships in the question, an information coverage checklist is constructed, and the integrity score of the candidate answer is calculated through the checklist. The candidate answers are identified using a professional dictionary in the power industry. The frequency and proportion of professional terms in the candidate answers are statistically analyzed, and the professionalism score of the candidate answers is calculated. The candidate answers are sorted using a heap sort algorithm. If the overall score of a candidate answer is higher than a preset threshold, the candidate answer with the highest overall score is selected as the final output result.

9. The method according to claim 1, characterized in that, Based on user feedback on the answers, a power industry dialogue strategy model is established using reinforcement learning algorithms. The dialogue strategy is adjusted according to this model to optimize the questioning methods and answer generation strategies for the next round. If the user is not satisfied with the answer, follow-up questions are generated based on the dialogue flow template to guide the user to provide additional details to obtain a complete slot filling result, including: A dialogue strategy model is constructed using the DeepQ-Network algorithm. The dialogue strategy model includes a state space and action space specific to the power industry. The state space includes user type, question category, and dialogue history; the action space includes question type, answer detail level, and use of technical terms. The dialogue strategy model is trained using a state-action-reward approach. The rewards are obtained from user feedback; Select the optimal action based on the Q-value output by the dialogue strategy model; The optimal action is used to determine the specific way to ask questions in the next round, including the wording, tone and complexity of the questions, and is also used to optimize the parameters for answer generation, including the level of detail of the answer and the frequency of use of technical terms. If the user is not satisfied with the answer, a follow-up question is generated using a sequence-to-sequence model based on an attention mechanism. The follow-up questions are selected and generated from a predefined follow-up question template library based on the current dialogue state and unfilled slots. Analyze user responses to follow-up questions using named entity recognition and dependency parsing techniques; The named entity recognition and dependency parsing technology is used to extract key information and fill it into the corresponding slots, update the dialogue state, and repeat the first two steps until a complete slot filling result is obtained or the preset maximum number of rounds is reached.

10. The method according to claim 1, characterized in that, If, after multiple rounds of dialogue, the system still cannot provide a satisfactory answer to the user, it will transfer the case to human customer service. The questions raised, the information obtained, and the attempted answers will be integrated into a service ticket, which will be followed up by human customer service. The human customer service process will be used as a new sample to optimize the dialogue knowledge base and dialogue strategy model of the power industry dialogue system, including: Receive service request information carrying a dialogue round count and user satisfaction evaluation index, the service request information being generated when a user initiates an electricity business inquiry; Based on the service request information, determine whether the number of dialogue rounds exceeds a preset threshold or whether the user satisfaction is continuously lower than a set value. If it exceeds the preset threshold or is lower than the set value, trigger the manual customer service transfer process. Named entity recognition and dependency parsing techniques are used to extract information from multi-turn dialogue content, resulting in user question descriptions and key information points; The user's problem description and key information points are processed using a text summarization algorithm to generate a structured service ticket; Based on the power service type and customer service expertise information in the service order, the service order management system automatically assigns the service order to the corresponding human customer service workstation. Record the human customer service process using real-time speech-to-text technology to obtain processing flow data; The processing flow data is analyzed using a long short-term memory network to extract key decision points and processing strategies. The knowledge base of the power industry dialogue system is updated based on the key decision points and processing strategies, and the dialogue strategy model is optimized.

Images