Power business interaction session processing method and system based on large language model

Abstract

The invention provides a power business interaction session processing method and system based on a large language model. The method comprises the following steps: firstly, obtaining a session semantic understanding vector sequence of a target power business interaction session record; and then semantic enhancement processing is carried out on the session semantic enhancement vector sequence to obtain a session semantic enhancement vector sequence. And finally, the reinforcement vector sequence is loaded into a first session record annotation network which is generated through network parameter learning according to a session semantic extraction instruction and a theme induction instruction, the loading data of the first session record annotation network is a first sample learning session record sequence, and the loading data of the first session record annotation network is a second sample learning session record sequence. The sequence comprises a first sample session semantic understanding vector sequence subjected to enhancement processing and subject induction distribution thereof. Through the mode, the session record annotation data of the target electric power business interaction session record can be generated, so that accurate analysis and processing of the electric power business interaction session are realized, the quality and efficiency of electric power services are improved, and the requirements of users on the aspect of electric power businesses are better met.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method and system for processing electric power business interactive conversations based on a large language model. Background Art

[0002] With the rapid development and intelligent transformation of the power industry, the types and complexity of power services are increasing, and the interactive conversations between users and power service providers are becoming more frequent and diverse. In the power business, accurately understanding and processing users' interactive conversations is crucial to providing high-quality services, solving problems, and optimizing business processes.

[0003] Traditional methods for processing interactive conversations in the power business usually rely on manual analysis or simple natural language processing technology, which has many limitations. Manual analysis is inefficient, difficult to handle large-scale conversation data, and easily affected by subjective factors, making it difficult to ensure the accuracy and consistency of the analysis results.

[0004] However, some existing natural language processing technologies may not be able to fully mine the deep semantic information in the conversation when processing power business interactive conversations, and the understanding of semantics is not accurate and comprehensive enough, which may lead to misjudgment of user needs and affect service quality and business decisions.

[0005] In addition, there is a lack of effective models and methods when labeling and classifying conversations, making it difficult to accurately extract the topics and key information of conversations, and unable to meet the growing needs of refined management and intelligent services in the power business.

[0006] Therefore, in order to improve the efficiency, accuracy and intelligence level of power business interactive conversation processing, it is necessary to develop an innovative method based on a large language model to achieve in-depth understanding and precise labeling of power business interactive conversations. Summary of the invention

[0007] In view of this, the purpose of the present application is to provide a method and system for processing electric power business interactive conversations based on a large language model.

[0008] According to a first aspect of the present application, a method for processing an electric power business interactive session based on a large language model is provided, the method comprising: Obtaining a conversation semantic understanding vector sequence of a target power business interaction conversation record; Performing semantic enhancement processing on the conversation semantic understanding vector sequence to generate a conversation semantic enhancement vector sequence; The conversation semantics enhancement vector sequence is loaded into a first conversation record annotation network to generate conversation record annotation data of the target power business interaction conversation record, wherein the first conversation record annotation network is a network parameter learning generation based on conversation semantics extraction instructions and topic induction instructions, the network loading data of the first conversation record annotation network is a first example learning conversation record sequence, the first example learning conversation record sequence includes a first example conversation semantics understanding vector sequence and a topic induction distribution of the first example conversation semantics understanding vector sequence, the topic induction distribution represents the conversation record annotation data of the first example conversation semantics understanding vector sequence, and the first example conversation semantics understanding vector sequence is a conversation semantics understanding vector that has been semantically enhanced.

[0009] According to the second aspect of the present application, a large language model-based electric power business interactive conversation processing system is provided, the large language model-based electric power business interactive conversation processing system comprises a machine-readable storage medium and a processor, the machine-readable storage medium stores machine-executable instructions, and when the processor executes the machine-executable instructions, the large language model-based electric power business interactive conversation processing system implements the aforementioned large language model-based electric power business interactive conversation processing method.

[0010] According to a third aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed, the aforementioned large language model-based power business interactive conversation processing method is implemented.

[0011] According to any one of the above aspects, the technical effect of the present application is: The embodiment of the present application can more deeply mine the semantic information in the conversation and reduce semantic ambiguity through semantic enhancement processing of the conversation semantic understanding vector sequence, so as to more accurately understand the needs and intentions of the user. The first conversation record annotation network is based on conversation semantic extraction instructions and topic induction instructions for learning, combined with the rich first example learning conversation record sequence, and can generate more accurate conversation record annotation data for the target power business interactive conversation, which helps to accurately classify and analyze the conversation. As a result, a large number of power business interactive conversation records can be processed quickly, adapting to conversation scenarios of different types and complexities, and providing strong support for the optimization and improvement of power business. That is, by accurately understanding and annotating conversations, it is possible to respond to and solve users' problems in power business in a timely and effective manner, improve user satisfaction, and promote the intelligent and efficient development of power services. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0013] Figure 1 A flowchart of a method for processing an interactive conversation of an electric power business based on a large language model is provided in an embodiment of the present application.

[0014] Figure 2 A schematic diagram of the component structure of a large language model-based electric power business interactive session processing system for implementing the above-mentioned large language model-based electric power business interactive session processing method provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical scheme and advantages of the embodiment of the present application clearer, the technical scheme in the embodiment of the present application will be clearly and completely described below according to the drawings in the embodiment of the present application. It should be understood that the drawings in the present application only serve the purpose of explanation and description, and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn in real proportion. The flowchart used in this application shows the operations implemented according to some embodiments of the embodiment of the present application. It should be understood that the operations of the flowchart can be implemented out of sequence, and the steps without logical context can be reversed in order or implemented simultaneously. In addition, those skilled in the art can add multiple other operations to the flowchart under the guidance of the content of the present application, and can also destroy multiple operations from the flowchart.

[0016] Figure 1 The flowchart of the method and system for processing electric power business interactive conversations based on a large language model provided by an embodiment of the present application is shown. It should be understood that in other embodiments, the order of some steps in the method for processing electric power business interactive conversations based on a large language model of this embodiment can be shared with each other according to actual needs, or some steps can be omitted or maintained. The detailed steps of the method for processing electric power business interactive conversations based on a large language model include: Step S110, obtaining a conversation semantic understanding vector sequence of a target power business interaction conversation record.

[0017] In detail, the target power business interactive conversation record refers to the actual interactive dialogue record about power business between the user and the customer service or system in the power system. The conversation semantic understanding vector sequence is obtained by converting each word or phrase in this target power business interactive conversation record into a vector of fixed length (i.e., word embedding), and encoding these vectors through natural language processing technology (such as recurrent neural network or long short-term memory network), thereby obtaining a vector sequence that can represent the semantics of the entire conversation.

[0018] In this embodiment, suppose that in the power system where the server is located, a user has an interactive conversation with a customer service staff about electricity bill inquiry. The user says: "I want to check this month's electricity bill, how do I do it?" The customer service staff replies: "You can check through our official website or mobile APP, just enter your user account and password." The server first performs text preprocessing on this target power business interactive conversation record, including noise removal and conversion to a unified encoding format. Then, the word embedding method in natural language processing technology is used to convert each word into a vector of fixed length. Next, these word vectors are encoded through models such as recurrent neural networks (RNN) or long short-term memory networks (LSTM), thereby obtaining a conversation semantic understanding vector sequence for the entire target power business interactive conversation record.

[0019] For example, after processing, the user's first sentence "I want to check this month's electricity bill, how do I do it?" may be represented as a sequence of 20 vectors [v1, v2, v3, ..., v20], each with a dimension of 100. The customer service staff's reply "You can query through our official website or mobile APP, just enter your user account and password." may be represented as a sequence of 30 vectors [v21, v22, v23, ..., v50], each with a dimension of 100.

[0020] Step S120: semantically enhance the conversation semantic understanding vector sequence to generate a conversation semantic enhancement vector sequence.

[0021] In detail, semantic enhancement processing is a method of further processing a conversation semantic understanding vector sequence, aiming to enhance or highlight key semantic information in the conversation and reduce semantic ambiguity. This can be achieved by using a pre-trained semantic enhancement processing model or performing masking processing (such as randomly selecting vectors for masking). The conversation semantic enhancement vector sequence is a conversation semantic understanding vector sequence after semantic enhancement processing, in which the vectors may contain richer semantic information, which helps to understand the conversation content more accurately.

[0022] In this embodiment, the server may perform semantic enhancement processing in a variety of ways.

[0023] Method 1: Using semantic enhancement processing model Assume that the server has a pre-trained semantic enhancement processing model. This model is self-learned based on a large number of second sample conversation semantic understanding vector sequences, sample conversation semantic enhancement vector sequences, and the first semantic vector generated by their fusion.

[0024] For the conversation semantic understanding vector sequence [v1, v2, v3, ..., v50] of the target power business interaction conversation record obtained above, the server loads it into the semantic reinforcement processing model. These vectors may be further analyzed and processed within the model, such as focusing on important semantic parts through the attention mechanism, or extracting more advanced semantic features through a multi-layer fully connected neural network. After processing, a conversation semantic reinforcement vector sequence [v'1, v'2, v'3, ..., v'50] is generated.

[0025] Method 2: Shielding The server can also directly perform masking on the conversation semantics understanding vector sequence to generate a conversation semantics reinforcement vector sequence. For example, some vectors in the sequence are randomly selected for masking, or masking is performed according to a certain rule (such as masking one vector every five vectors). The masked vectors can be replaced by zero vectors or other preset special vectors. After the masking process, the conversation semantics reinforcement vector sequence is obtained.

[0026] Step S130: Load the conversation semantics enhancement vector sequence into a first conversation record annotation network to generate conversation record annotation data of the target power business interaction conversation record, wherein the first conversation record annotation network is generated by network parameter learning based on conversation semantics extraction instructions and topic induction instructions, and the network loading data of the first conversation record annotation network is a first example learning conversation record sequence, and the first example learning conversation record sequence includes a first example conversation semantics understanding vector sequence and a topic induction distribution of the first example conversation semantics understanding vector sequence, and the topic induction distribution represents the conversation record annotation data of the first example conversation semantics understanding vector sequence, and the first example conversation semantics understanding vector sequence is a conversation semantics understanding vector that has been processed with semantic enhancement.

[0027] In detail, the first conversation record annotation network is a neural network model that performs network parameter learning based on conversation semantic extraction instructions and topic summarization instructions. Its function is to classify and annotate conversation semantic vector sequences that have undergone semantic enhancement processing to generate conversation record annotation data.

[0028] The conversation semantic extraction instruction is an instruction or task that requires the network model to extract key semantic information from the conversation. This is usually achieved by training the model to identify keywords, phrases or sentences in the conversation. The topic induction instruction is an instruction or task that requires the network model to classify the conversation into a specific topic or category. For example, electricity bill inquiry, fault repair, application for new meter installation, etc. This is usually achieved by training the model to identify the overall topic or intention of the conversation.

[0029] The network loading data refers to data used for training or loading into a neural network model. In the first session record annotation network, the network loading data is a first example learning session record sequence.

[0030] The first example learning conversation record sequence is a set of sample data for training the first conversation record annotation network, including the first example conversation semantic understanding vector sequence and the topic induction distribution corresponding to these vector sequences. The first example conversation semantic understanding vector sequence refers to the sample data of the conversation semantic understanding vector sequence after semantic enhancement processing, which is used to train the first conversation record annotation network.

[0031] The topic summary distribution is a vector or probability distribution, which indicates the probability that the session record belongs to different topics or categories. For example, [0.9, 0.05, 0.05] indicates that there is a 90% probability that the session belongs to the electricity fee inquiry category.

[0032] That is, in this embodiment, the first conversation record annotation network is generated by learning network parameters based on the conversation semantic extraction instruction and the topic induction instruction. Its network loading data is the first example learning conversation record sequence, and the first example learning conversation record sequence includes the first example conversation semantic understanding vector sequence and its topic induction distribution.

[0033] Assume that the first example learning conversation record sequence contains various conversations about electricity business, such as electricity bill inquiry, fault repair, new electricity meter application, etc., and each conversation has a corresponding topic summary distribution. For example, the topic summary distribution of "electricity bill inquiry" may be [1, 0, 0], indicating that it belongs to the electricity bill inquiry category; the topic summary distribution of "fault repair" may be [0, 1, 0], indicating that it belongs to the fault repair category; the topic summary distribution of "new electricity meter application" may be [0, 0, 1], indicating that it belongs to the new electricity meter application category.

[0034] The server loads the conversation semantics reinforcement vector sequence of the target power business interaction conversation record into the first conversation record annotation network. The neurons in the network will calculate and process these vectors, and finally output a predicted topic summary distribution through a multi-layer convolutional neural network (CNN) or Transformer architecture.

[0035] For example, for the electricity bill inquiry session mentioned above, after being processed by the first session record annotation network, the output topic summary distribution may be [0.9, 0.05, 0.05], indicating that there is a 90% probability that the session belongs to the electricity bill inquiry category, a 5% probability that it belongs to the fault repair category, and a 5% probability that it belongs to the new meter application category.

[0036] The server summarizes the distribution of the output topics and determines the session record annotation data of the target power business interaction session record, that is, it is determined that it mainly belongs to the electricity fee inquiry category.

[0037] Suppose in another scenario, the conversation between the user and the customer service is about reporting a power failure. The user said: "My home suddenly lost power and I don't know why." The customer service replied: "Please check whether it is a circuit breaker. If not, we will arrange for maintenance personnel to come to check as soon as possible." The server processes this conversation according to the above step S110 to obtain a conversation semantic understanding vector sequence. Then, in step S120, a conversation semantic enhancement vector sequence is generated through a semantic enhancement processing model or a shielding process. Finally, in step S130, the conversation is loaded into the first conversation record annotation network, and the output topic summary distribution may be [0.02, 0.95, 0.03], thereby determining that the conversation belongs to the fault repair category.

[0038] For example, the conversation between the user and the customer service is about applying for a new electricity meter. The user said: "I just bought a house and want to apply for a new electricity meter." The customer service replied: "OK, please prepare the relevant property certificate and ID card, and we will process it for you." The server also processes steps S110 to S130, and the final output topic summary distribution may be [0.01, 0.02, 0.97], determining that the session belongs to the category of application for new electricity meter installation.

[0039] This design can effectively obtain the conversation semantic understanding vector sequence of the target power business interaction conversation record, perform semantic enhancement processing, and use the first conversation record annotation network to generate accurate conversation record annotation data, thereby realizing intelligent analysis and processing of the power business interaction conversation.

[0040] Based on the above steps, the embodiment of the present application can more deeply mine the semantic information in the conversation and reduce semantic ambiguity through semantic enhancement processing of the conversation semantic understanding vector sequence, so as to more accurately understand the needs and intentions of the user. The first conversation record annotation network is based on conversation semantic extraction instructions and topic induction instructions for learning, combined with the rich first example learning conversation record sequence, it can generate more accurate conversation record annotation data for the target power business interactive conversation, which is helpful for accurate classification and analysis of the conversation. As a result, a large number of electric power business interactive conversation records can be processed quickly, adapting to conversation scenarios of different types and complexities, and providing strong support for the optimization and improvement of the electric power business. That is, by accurately understanding and annotating the conversation, it is possible to respond to and solve users' problems in the electric power business in a timely and effective manner, improve user satisfaction, and promote the intelligent and efficient development of electric power services.

[0041] In a possible implementation, step S120 includes: Step S121: Load the conversation semantic understanding vector sequence into the semantic enhancement processing model to generate the conversation semantic enhancement vector sequence. The semantic enhancement processing model performs self-learning generation based on the second sample conversation semantic understanding vector sequence, the sample conversation semantic enhancement vector sequence and the first semantic vector. The first semantic vector is generated by the fusion of the second sample conversation semantic understanding vector sequence and the sample conversation semantic enhancement vector sequence.

[0042] Alternatively, in step S122, the conversation semantics understanding vector sequence is masked to generate the conversation semantics enhancement vector sequence.

[0043] In this embodiment, in the power system, the server is responsible for processing a large number of power business interaction session records.

[0044] Assume that the server receives a target power business interactive session record, such as a discussion between a user and customer service about abnormal power consumption. The user says, "I feel that my power consumption has suddenly increased a lot this month. Is there something wrong with the meter?" The customer service replies, "Hello, we will arrange for staff to check it. It may be that some electrical appliances are leaking electricity or your power usage habits have changed." The server first performs text preprocessing and word embedding operations on the conversation according to the normal processing flow, and encodes it through models such as recurrent neural networks or long short-term memory networks to obtain a conversation semantic understanding vector sequence, assumed to be [v1, v2, v3, ..., v40].

[0045] Method 1: Using semantic enhancement processing model The server has a pre-trained semantic enhancement processing model. This semantic enhancement processing model is self-learned based on a large number of second sample conversation semantic understanding vector sequences, sample conversation semantic enhancement vector sequences, and the first semantic vectors generated by their fusion.

[0046] Assume that a large number of second example conversation semantic understanding vector sequences are stored in the database of the server, for example: - Conversations about electricity bill disputes: [v101, v102, v103, ..., v150] - Conversations about power outage notifications: [v201, v202, v203, ..., v250] There are also corresponding sample conversation semantics reinforcement vector sequences, such as: - Hardened electricity bill dispute sessions: [v151, v152, v153, ..., v200] - Hardened outage notification sessions: [v251, v252, v253, ..., v300] The second sample conversation semantic understanding vector sequence and the sample conversation semantic enhancement vector sequence are fused to generate a first semantic vector, for example, by weighted averaging or concatenation.

[0047] For the conversation semantic understanding vector sequence [v1, v2, v3, ..., v40] of the currently acquired target power business interaction conversation record, the server loads it into the semantic reinforcement processing model. The importance of each vector may be analyzed through the attention mechanism within the model.

[0048] For example, in this discussion about abnormal electricity consumption, the vectors corresponding to keywords such as "sudden increase in electricity consumption" and "problem with the electricity meter" may be given higher weights. Then, a multi-layer fully connected neural network is used to extract deeper semantic features.

[0049] After processing, a sequence of conversation semantics enhanced vectors is generated, assuming it is [v'1, v'2, v'3, ..., v'40]. These enhanced vectors may have changed values ​​in some dimensions or added new semantic information to better reflect the core semantics of the conversation.

[0050] Method 2: Shielding The server may also choose to directly perform masking processing on the session semantics understanding vector sequence to generate a session semantics enhancement vector sequence.

[0051] Assume that the server sets a blocking rule of blocking every three vectors. For the session semantic understanding vector sequence [v1, v2, v3, ..., v40], then the vectors v4, v7, v10, etc. are blocked.

[0052] The masked vectors are replaced by zero vectors. After this processing, a sequence of conversation semantics reinforcement vectors is obtained, such as [v1, v2, v3, 0, v5, v6, v7, 0, ...].

[0053] Through this masking process, the model may pay more attention to the key information carried by the unmasked vectors in subsequent learning and analysis, thereby improving the ability to capture important semantics.

[0054] For example, another target power business interactive session record is about power equipment maintenance. The user said: "The transformer seems to be faulty and keeps making noises." The customer service replied: "We will send someone to check and repair it as soon as possible." The server also first obtains the conversation semantic understanding vector sequence, assumed to be [v51, v52, v53, ..., v70].

[0055] When using the semantic reinforcement processing model, the model may analyze and reinforce the current vector sequence based on previously learned sample vector sequences and reinforcement vector sequences related to similar maintenance, highlighting key information such as "transformer failure" and "ringing".

[0056] When performing the shielding process, if the rule of shielding every two vectors is followed, vectors such as v53, v55, and v57 are shielded to generate a corresponding sequence of conversation semantics reinforcement vectors.

[0057] Through the above different methods of semantic enhancement processing, the server can provide more representative and effective input data for the subsequent conversation record annotation network, thereby improving the analysis and annotation accuracy of power business interactive conversations.

[0058] Suppose there is another conversation about the progress of the power project. The user said: "I want to know when the power project in our community will be completed?" The customer service replied: "It is expected to be completed in about a month, depending on the progress of the construction." After the server obtains its session semantic understanding vector sequence [v81, v82, v83, ..., v100], if a semantic reinforcement processing model is adopted, the model may strengthen the semantic vectors related to time and construction based on sample data of similar construction progress inquiries in the past.

[0059] If shielding processing is adopted, for example, every 4 vectors are shielded, a new sequence of conversation semantics reinforcement vectors is generated.

[0060] In this way, when processing various types of power business interactive sessions, the server can flexibly use these two semantic enhancement processing methods to adapt to different session characteristics and needs, laying the foundation for accurate annotation of session records.

[0061] Let's assume that there is another conversation about electricity complaints. The user said emotionally: "Your service is too bad. I have been waiting for a long time and no one has come to solve the problem!" The customer service hurriedly apologized and said that they would handle it as soon as possible.

[0062] After the server obtains the conversation semantic understanding vector sequence, it can focus on strengthening the user's dissatisfaction and the core issues of the complaint through the semantic reinforcement processing model.

[0063] Or through masking processing, some less critical information can be selectively ignored to highlight the important semantic parts.

[0064] In short, no matter what type of electric business interactive session, the server can select the appropriate semantic enhancement processing method according to the specific situation, generate a session semantic enhancement vector sequence, and prepare for subsequent analysis and processing.

[0065] Suppose there is another more complex conversation involving a detailed discussion of an electricity contract. The user and customer service have a detailed exchange of views on some terms.

[0066] After obtaining the conversation semantic understanding vector sequence, the server uses the semantic enhancement processing model to comprehensively consider the large amount of contract-related sample data accumulated previously and extract more accurate semantic features.

[0067] Through shielding processing, data can be simplified and key information can be more prominent.

[0068] In the actual operation of the power system, a large number of various conversation records are generated every day. By continuously applying these semantic enhancement processing methods, the server can gradually improve its ability to understand and analyze conversations, providing strong support for the optimization and improvement of power business.

[0069] For example, for a conversation about consulting on preferential electricity policies, the server can accurately enhance the relevant semantics and provide more valuable information for subsequent judgment of the topic and category of the conversation.

[0070] For example, when handling a conversation about emergency handling of power failures, the server can quickly identify key issues and arrange corresponding handling measures in a timely manner through effective semantic enhancement.

[0071] As time goes by, the sample data accumulated by the server becomes richer, the semantic enhancement processing model is continuously optimized, and the shielding processing rules become more reasonable, making the processing of power business interactive sessions more efficient and accurate.

[0072] Assume that in a certain period of time, users mainly inquire about the adjustment of electricity prices. By processing the semantics of these conversations, the server can better summarize and generalize users' concerns and needs, and provide a reference for the power department to formulate relevant policies.

[0073] Assuming that there is a discussion session about the connection of renewable energy to the power grid, the server can also extract key information through the above semantic enhancement processing method to provide valuable suggestions for the planning and development of the power system.

[0074] Whether it is daily consultation from ordinary users or communication on large-scale electricity demand from corporate users, the server can ensure the smooth interaction of power business and improve service quality and user satisfaction through precise semantic enhancement processing.

[0075] Assume that during a peak power consumption period, a large number of users raise questions and suggestions about power restrictions and allocation. The server can sort out and summarize this information in a timely manner through fast and accurate semantic enhancement processing, providing strong data support for power dispatch decision-making.

[0076] For example, when the power system undergoes technical upgrades and renovations, users ask about the possible impacts and changes. Through effective semantic enhancement processing, the server can clearly understand the users' concerns and expectations, and thus better communicate and explain to the users.

[0077] This design can flexibly use semantic enhancement processing methods to deal with various complex power business interactive conversation situations, providing solid technical support for the stable operation of the power system and high-quality services.

[0078] In a possible implementation, the method further includes: Step A110, obtaining a second sample learning conversation record sequence and a basic semantic reinforcement processing model, wherein the second sample learning conversation record sequence includes the second sample conversation semantic understanding vector sequence, the sample conversation semantic reinforcement vector sequence and the first semantic vector.

[0079] Step A120, loading the second sample learning conversation record sequence into the basic semantic reinforcement processing model for semantic reinforcement to generate a second semantic vector, and the second semantic vector is generated by hybrid simulation based on the second sample conversation semantic understanding vector sequence and the sample conversation semantic reinforcement vector sequence.

[0080] Step A130: calculating a first network learning cost according to the second semantic vector and the first semantic vector.

[0081] Step A140: Training the basic semantic enhancement processing model based on the first network learning cost to generate a semantic enhancement processing model.

[0082] First, the server obtains the second sample learning session record sequence, assuming that it includes multiple different types of power business session records. For example: -A conversation about electricity bill arrears notification, whose second example conversation semantic understanding vector sequence is [v1001, v1002, v1003, ..., v1050].

[0083] -A conversation about the arrangement of power equipment upgrades, whose sample conversation semantic reinforcement vector sequence is [v2001, v2002, v2003, ..., v2050].

[0084] At the same time, a first semantic vector is generated by a fusion operation on the two sequences, for example, by a specific weighted fusion method.

[0085] The server has an initial basic semantic enhancement processing model, and loads the acquired second sample learning session record sequence into the basic semantic enhancement processing model.

[0086] The basic semantic enhancement processing model performs hybrid simulation processing on the input second sample conversation semantic understanding vector sequence and the sample conversation semantic enhancement vector sequence according to the preset algorithm and parameters. For example, the second semantic vector is generated by interactively calculating the vector elements in the two sequences, performing feature fusion and other operations.

[0087] Assume that for the conversation about the above-mentioned electricity bill arrears notification, the second semantic vector generated is [v1101, v1102, v1103, ..., v1150]; for the conversation about the arrangement of power equipment upgrade, the second semantic vector generated is [v2101, v2102, v2103, ..., v2150].

[0088] Next, the server calculates the learning cost of the first network based on the generated second semantic vector and the first semantic vector obtained by fusion. The calculation method may be to compare the differences between the two semantic vectors in various dimensions and use cost functions such as mean square error and cross entropy to quantify the differences.

[0089] For example, for two vectors of electricity bill arrears notification sessions, the sum of the squares of the differences between their corresponding elements is calculated and averaged by dividing by the vector length.

[0090] Based on the calculated first network learning cost, the server trains the basic semantic reinforcement processing model and adjusts the weight parameters in the model so that when the model processes a new sample learning session record, the second semantic vector generated can be closer to the first semantic vector, that is, more consistent with the expected semantic representation.

[0091] In another scenario, assume that the second sample learning conversation record sequence contains a conversation about emergency repair of power failure, and its second sample conversation semantic understanding vector sequence is [v3001, v3002, v3003, ..., v3050], and the sample conversation semantic reinforcement vector sequence is [v4001, v4002, v4003, ..., v4050].

[0092] After the fusion simulation processing of the basic semantic enhancement processing model, the generated second semantic vector is [v3101, v3102, v3103, ..., v3150].

[0093] Then, the first network learning cost is calculated by comparing with the first semantic vector generated by fusion, and the model is further trained and optimized based on the cost.

[0094] In this way, by continuously acquiring new second sample learning conversation record sequences and repeating the above process, the server's basic semantic enhancement processing model is gradually trained to become a mature model that can accurately perform semantic enhancement processing, providing more effective support for subsequent processing of various power business interactive conversations.

[0095] For example, we obtained a conversation record about consulting on opening an electricity account for a new user and processed and trained it according to the same process to continuously improve the performance and accuracy of the model.

[0096] In actual power business operations, such operations are continuously performed so that the model can adapt to the ever-changing and rich power business session scenarios, providing strong guarantees for the efficient operation and high-quality services of the power system.

[0097] In a possible implementation, step A110 includes: Step A111, obtaining a first sample learning session record, a sample semantic enhancement session record and a first electric power business interaction session record, wherein the first electric power business interaction session record is generated by fusion of the first sample learning session record and the sample semantic enhancement session record.

[0098] Step A112, in non-online mode, semantic extraction is performed on the first sample learning session record to generate the second sample session semantic understanding vector sequence, and semantic extraction is performed on the sample semantic enhancement session record to generate the sample session semantic enhancement vector sequence, and semantic extraction is performed on the first power business interaction session record to generate the first semantic vector.

[0099] In this embodiment, first, the server obtains the first example learning session record. For example, a session record about a user applying for increased power capacity, the user said: "I have added a lot of new electrical appliances at home, the existing power capacity is not enough, I want to apply for an increase." At the same time, the server obtains sample semantically enhanced conversation records. For example, by replacing keywords in the first sample learning conversation record, the sample semantically enhanced conversation record replaces "power capacity" with "power consumption", which becomes "My family has added a lot of new electrical appliances, and the existing power consumption is not enough. I want to apply for an increase." Then, the first example learning session record and the example semantic enhancement session record are fused to generate the first power business interaction session record.

[0100] In the offline mode, the server performs semantic extraction on the first sample learning session record. For example, the session record of the above application for capacity expansion is processed using a word vector model and a neural network, and converted into a series of digital vectors, thereby generating a second sample session semantic understanding vector sequence [v1, v2, v3, ..., vn].

[0101] Next, the same semantic extraction operation is performed on the sample semantically enhanced conversation record to generate a sample conversation semantically enhanced vector sequence [v'1, v'2, v'3, ..., v'm].

[0102] Finally, semantic extraction is performed on the fused first electric business interaction session record to generate a first semantic vector [v''1, v''2, v''3, ..., v''k].

[0103] For example, in another scenario, the first sample learning session record is about a user asking about the tiered electricity price policy, "I'm not sure how the tiered electricity prices here are divided." The sample semantic enhancement session record is obtained by adjusting its sequence, "I'm not sure how the tiered electricity prices here are divided." After the server merges the two conversation records, it performs semantic extraction respectively in the offline mode to generate the corresponding second sample conversation semantic understanding vector sequence, the sample conversation semantic reinforcement vector sequence and the first semantic vector.

[0104] For example, the first example learning session record is a user feedback about abnormal meter readings, "The meter readings this month seem to be wrong, much higher than usual." The sample semantic enhancement session record is obtained by generating semantic random perturbations, such as "The meter readings this month seem to be wrong, much higher than usual." After fusion, semantic extraction is performed to obtain the required vector sequence and semantic vector.

[0105] By continuously acquiring and processing various first-example learning session records and example semantic enhancement session records, we can enrich data samples and provide strong support for subsequent model training and optimization.

[0106] In a possible implementation, step A111 includes: The first example learning session record is obtained, and a semantic random perturbation session record is generated, wherein the semantic random perturbation session record is used as the example semantic reinforcement session record. The semantic random perturbation session record is merged with the first example learning session record to generate the first power business interaction session record.

[0107] Alternatively, the first sample learning session record is obtained, and the sequence of the first sample learning session record is adjusted to generate a rearranged power business interaction session record of the first sample learning session record, and the rearranged power business interaction session record is used as the sample semantic enhancement session record. The rearranged power business interaction session record is merged with the first sample learning session record to generate the first power business interaction session record.

[0108] Alternatively, the first example learning session record is obtained, and the first example learning session record is subjected to local attention weight content screening to generate a screened power business interaction session record of the first example learning session record, and the screened power business interaction session record is used as the example semantic reinforcement session record. The screened power business interaction session record is merged with the first example learning session record to generate the first power business interaction session record.

[0109] Alternatively, the first example learning session record is obtained, and the keywords of the first example learning session record are replaced to generate a keyword replacement session record of the first example learning session record, and the keyword replacement session record is used as the example semantic enhancement session record. The keyword replacement session record is merged with the first example learning session record to generate the first power business interaction session record.

[0110] In the power system, the server needs to obtain and process various sample learning session records for subsequent analysis and model training.

[0111] First case: The server first obtains a first-instance learning session record, for example, the user says: "Why is my electricity bill so high this month? Is there something wrong with the meter?" Then, the server generates a semantically random perturbation session record by randomly changing some words or sentence structures, such as: "Why is my electricity bill so high this month? Is there something wrong with the meter?" This semantically random perturbation session record is used as a sample semantically enhanced session record. Finally, the semantically random perturbation session record is fused with the original first-instance learning session record to generate the first power business interaction session record.

[0112] Second case: The server obtains the first example learning session record as: "I want to install a new electricity meter. What materials do I need to prepare?" Then, the sequence of this session record is adjusted to generate a rearranged electricity business interaction session record, such as: "What materials do I need to prepare? I want to install a new electricity meter?" This rearranged electricity business interaction session record is used as a sample semantic enhancement session record, and is merged with the original first example learning session record to obtain the first electricity business interaction session record.

[0113] The third case: The server obtains the first example learning session record: "The community has been experiencing power outages recently, affecting normal life." Then, it performs local attention weight content screening, such as screening out the key part "The community has been experiencing power outages" to generate a screened power business interaction session record: "The community has been experiencing power outages." This screened record is used as a sample semantic reinforcement session record and merged with the first example learning session record to form the first power business interaction session record.

[0114] The fourth case: The server obtains the first example learning session record as: "The process of electricity repair reporting is too complicated, can it be simplified?" Then, it replaces the keywords therein, such as replacing "electricity repair reporting" with "electricity maintenance", and generates a keyword replacement session record: "The process of electricity maintenance is too complicated, can it be simplified?" This record is used as a sample semantic enhancement session record and is merged with the original first example learning session record to obtain the first electricity business interaction session record.

[0115] Through the above different methods, the server obtains rich and diverse sample data, providing more comprehensive and effective data support for subsequent power business session analysis and processing.

[0116] In a possible implementation, the method further includes: Step B110, obtaining a first sample learning conversation record sequence and a basic conversation record annotation network, wherein the second sample learning conversation record sequence includes the first sample conversation semantic understanding vector sequence and the topic induction distribution of the first sample conversation semantic understanding vector sequence, the basic conversation record annotation network includes a basic neural network layer, a basic topic classification function layer and at least one basic semantic extraction function layer, the topic induction distribution represents the conversation record annotation data of the first sample conversation semantic understanding vector sequence, the first sample conversation semantic understanding vector sequence is a conversation semantic understanding vector that has been semantically enhanced, the processing data of the at least one basic semantic extraction function layer is used to train the basic neural network layer and the basic topic classification function layer, the processing data of the basic topic classification function layer is the processing data of the basic conversation record annotation network, and the first sample conversation semantic understanding vector sequence represents the continuous semantic context vector of the electric power business interaction conversation record.

[0117] Step B120: Load the first sample learning conversation record sequence into the basic conversation record annotation network, and perform network parameter learning based on the topic induction instruction and the conversation semantic extraction instruction to generate the first conversation record annotation network.

[0118] In this embodiment, first, the server obtains the first example learning session record sequence. For example, there is a series of session records: -Session 1: The user said, "I want to check the details of last month's electricity bill. How can I check it?" The customer service replied, "You can enter the meter number through the mobile APP or the online business hall to check." After processing, the first example of this session's session semantic understanding vector sequence may be [v101, v102, v103, ..., v150], and its topic summary distribution is [1, 0, 0], indicating that it belongs to the electricity bill query category.

[0119] -Session 2: The user said, "The electric pole near my home seems to have fallen down. It's very dangerous." The customer service responded, "We will send someone to check it out right away." The first example of this session's session semantic understanding vector sequence might be [v201, v202, v203, ..., v250], and the topic summary distribution is [0, 1, 0], indicating that it belongs to the category of power equipment failure.

[0120] At the same time, the server also obtains the basic conversation record annotation network. This network includes a basic neural network layer, a basic topic classification function layer, and multiple basic semantic extraction function layers. The basic neural network layer is used to perform preliminary processing and feature extraction on the input vector, the basic topic classification function layer is used to classify and judge the topic of the conversation, and multiple basic semantic extraction function layers are used to extract deeper semantic information from the vector.

[0121] Next, the server loads the first example learning session record sequence into the basic session record annotation network.

[0122] For session one, the server loads its first example session semantic understanding vector sequence [v101, v102, v103, ..., v150] into the basic neural network layer and the basic topic classification function layer. The basic neural network layer performs operations such as convolution and pooling on these vectors to extract preliminary features. The basic topic classification function layer generates session topic classification prediction data based on these features. Assuming the predicted data is [0.8, 0.1, 0.1], the server then calculates the learning cost of the second network based on this predicted data and the known topic summary distribution [1, 0, 0]. If the predicted data is far from the actual topic summary distribution, such as in this example, the predicted probability of the electricity bill query category is 0.8 instead of 1, then the learning cost will be relatively high.

[0123] Based on the calculated learning cost of the second network, the server iteratively trains the basic neural network layer and the basic subject classification function layer, adjusts the parameters of the network, and generates a more accurate first neural network layer and first subject classification function layer.

[0124] Then, the server randomly selects 5 vectors (assuming v101-v105) from the first example conversation semantic understanding vector sequence of conversation one, and fuses these 5 vectors according to the conversation logic order to generate the first conversation semantic understanding vector to be learned. This first conversation semantic understanding vector to be learned is loaded into the first neural network layer and the first basic semantic extraction function layer to generate the first conversation semantic estimation vector. The third network learning cost is calculated based on this first conversation semantic estimation vector and the sixth vector (i.e., v106) in the first example conversation semantic understanding vector sequence.

[0125] Based on the third network learning cost, the first neural network layer and the first subject classification function layer are trained again to generate the second neural network layer and the second subject classification function layer.

[0126] Afterwards, the server randomly determines three semantic feature nodes (assuming they are v110, v120, and v130) from the first example session semantic understanding vector sequence of session 1 and performs masking processing to generate a second session semantic understanding vector to be learned. The second session semantic understanding vector to be learned is loaded into the second neural network layer and the second basic semantic extraction function layer to generate a second session semantic estimation vector. The fourth network learning cost is calculated based on the second session semantic estimation vector and the actual feature vectors of the three masked semantic feature nodes (v110, v120, and v130).

[0127] The second neural network layer and the second topic classification function layer are trained based on the fourth network learning cost, the performance of the network is gradually optimized, and finally the first conversation record annotation network is generated.

[0128] For the second session, the server follows the same process for processing and training. The first example session semantic understanding vector sequence [v201, v202, v203, ..., v250] is loaded into the network, the session topic classification prediction data is generated, the second network learning cost is calculated and the corresponding training is performed. Then the vectors are selected and fused, the third network learning cost is calculated, the shielding process is performed, the fourth network learning cost is calculated, and the network parameters are continuously optimized to generate the first session record annotation network that can accurately annotate the power equipment fault session.

[0129] For example, in another conversation record, the user said, "I want to apply for time-of-use electricity price service, what procedures are required?" The customer service replied, "You need to provide your ID card, real estate certificate and a recent electricity bill." The first example conversation semantic understanding vector sequence of this conversation is assumed to be [v301, v302, v303, ..., v350], and the topic summary distribution is [0, 0, 1], indicating that it belongs to the electricity price service processing category.

[0130] The server also loads it into the basic conversation record annotation network, processes and trains it according to the above steps, and continuously adjusts the network parameters so that the network can accurately identify and annotate the topics of such conversations.

[0131] By repeatedly loading, calculating learning costs and training multiple such first-example learning session record sequences, the server continuously optimizes the parameters of the basic session record annotation network to generate a first session record annotation network that can accurately understand and annotate various power business interactive sessions.

[0132] In the actual operation of power business, a large number of new session records are generated every day. With the generated first session record annotation network, the server can quickly and accurately analyze and annotate these new sessions, providing strong support for the optimization and improvement of power services.

[0133] For example, when there is a new user session "I have questions about this month's electricity bill, can you help me explain it?" The server can quickly process it into a sequence of session semantic understanding vectors and load it into the trained first session record annotation network, output accurate topic summary distribution, and determine that it belongs to the electricity bill query category, thereby providing clear guidance for subsequent services.

[0134] For example, the new conversation is "Our factory wants to expand its power capacity, how do we apply?" The server can also accurately determine that it belongs to the power business application category through the first conversation record annotation network and take corresponding processing measures.

[0135] As time goes by, the server can continuously obtain new first-example learning session record sequences, update and optimize the first session record annotation network, so that it can always adapt to the development and changes of the power business and provide users with better quality and more efficient services.

[0136] Assume that during the peak period of electricity consumption in summer, a large number of users' conversations involve issues such as increased electricity bills and electricity restrictions. The server can quickly classify and annotate these conversations through the first conversation record annotation network, timely summarize the user's focus and needs, and provide data support for the power department to formulate reasonable response strategies.

[0137] For another example, when the power sector launches a new preferential policy, the number of consultation sessions of users increases. The server uses the first session record annotation network to accurately identify the topics of these sessions and provide users with accurate and detailed policy interpretation and handling guidance.

[0138] In summary, by acquiring the first example learning session record sequence and the basic session record annotation network, and performing network parameter learning to generate the first session record annotation network, the server can realize intelligent and efficient processing of power business interactive sessions, and improve the quality of power services and user satisfaction.

[0139] In a possible implementation, step B120 includes: Step B121: Load the first sample conversation semantic understanding vector sequence into the basic neural network layer and the basic topic classification function layer to generate conversation topic classification prediction data.

[0140] Step B122, calculating a second network learning cost based on the conversation topic classification prediction data and the topic summary distribution, and iteratively training the basic neural network layer and the basic topic classification function layer based on the second network learning cost to generate a first neural network layer and a first topic classification function layer.

[0141] Step B123: arbitrarily select X sample conversation semantic understanding vectors from the first sample conversation semantic understanding vector sequence, and fuse the X sample conversation semantic understanding vectors according to the conversation logic order to generate a first conversation semantic understanding vector to be learned, where X is a positive integer.

[0142] Step B124: Load the first to-be-learned conversation semantic understanding vector into the first neural network layer and the first basic semantic extraction function layer in the at least one basic semantic extraction function layer to generate a first conversation semantic estimation vector.

[0143] Step B125: Calculate a third network learning cost based on the first conversation semantics estimation vector and the X+1th sample conversation semantics understanding vector in the first sample conversation semantics understanding vector sequence.

[0144] Step B126: Training the first neural network layer and the first topic classification function layer based on the third network learning cost to generate a second neural network layer and a second topic classification function layer.

[0145] Step B127: randomly determine a set number of first semantic feature nodes from the first sample conversation semantic understanding vector sequence, perform masking processing on them, and generate a second conversation semantic understanding vector to be learned.

[0146] Step B128: Load the second to-be-learned conversation semantic understanding vector into the second neural network layer and the second basic semantic extraction function layer in the at least one basic semantic extraction function layer to generate a second conversation semantic estimation vector.

[0147] Step B129: Calculate the fourth network learning cost based on the second conversation semantic estimation vector and the actual feature vector of the first semantic feature node, train the second neural network layer and the second topic classification function layer based on the fourth network learning cost, and generate the first conversation record annotation network.

[0148] First, the server obtains a series of first-instance learning session record sequences. For example, there is a session record like this: the user says, "My electricity bill this month is too high. Is the meter reading wrong?" The customer service replies, "We will arrange for someone to check. Don't worry." After processing, the first-instance session semantic understanding vector sequence of this session is [v1, v2, v3, ..., v50].

[0149] The server loads this first example conversation semantic understanding vector sequence into the basic neural network layer and basic topic classification function layer of the basic conversation record annotation network. The basic neural network layer performs a series of calculations and processing on these vectors, and the basic topic classification function layer generates conversation topic classification prediction data based on the processing results. Assume that the generated prediction data is [0.7, 0.2, 0.1], which means that there is a 70% probability that it is related to electricity bill issues, a 20% probability that it is other categories, and a 10% probability that it is another category.

[0150] Then, the server compares the conversation topic classification prediction data with the known topic summary distribution (such as [1, 0, 0], indicating that it is definitely the electricity bill problem category) and calculates the learning cost of the second network. If the predicted data is far from the actual topic summary distribution, the learning cost will be higher.

[0151] Based on the calculated learning cost of the second network, the server iteratively trains the basic neural network layer and the basic subject classification function layer, and generates a more optimized first neural network layer and first subject classification function layer by adjusting the weights and parameters in the network.

[0152] Next, the server randomly selects five vectors (assuming v1-v5) from the first example conversation semantic understanding vector sequence, fuses these five vectors according to the logical order of the conversation, and generates the first conversation semantic understanding vector to be learned. Then, the first conversation semantic understanding vector to be learned is loaded into the first neural network layer and the first basic semantic extraction function layer, and the first conversation semantic estimation vector is generated after processing.

[0153] Afterwards, the server calculates the third network learning cost based on the first conversation semantics estimation vector and the sixth vector (ie, v6) in the first example conversation semantics understanding vector sequence.

[0154] According to the learning cost of the third network, the first neural network layer and the first subject classification function layer are trained again and further optimized to generate the second neural network layer and the second subject classification function layer.

[0155] Subsequently, the server randomly determines three semantic feature nodes (assumed to be v10, v20, and v30) from the first example conversation semantic understanding vector sequence and performs masking processing to generate a second conversation semantic understanding vector to be learned.

[0156] The second conversation semantics understanding vector to be learned is loaded into the second neural network layer and the second basic semantics extraction function layer to generate a second conversation semantics estimation vector.

[0157] Next, the fourth network learning cost is calculated based on the second session semantic estimation vector and the actual feature vectors of the three masked semantic feature nodes (v10, v20, v30).

[0158] Finally, the second neural network layer and the second topic classification function layer are trained based on the fourth network learning cost, and the network parameters are continuously optimized and adjusted to finally generate a first conversation record annotation network that can accurately annotate and understand such conversations.

[0159] For example, in another conversation record, the user said, "I want to apply for a new electricity meter. What materials do I need to prepare?" The customer service replied, "You need to bring your ID card, property certificate, and application form." The first example conversation semantic understanding vector sequence is [v51, v52, v53, ..., v80].

[0160] The server also processes according to the above steps. First, it is loaded into the basic neural network layer and the basic topic classification function layer to generate the conversation topic classification prediction data, and then the second network learning cost is calculated and the corresponding iterative training is performed. Then 5 vectors are selected for fusion to generate the first conversation semantic understanding vector to be learned, and the third network learning cost is calculated and trained. Then 3 semantic feature nodes are randomly masked to generate the second conversation semantic understanding vector to be learned, and the fourth network learning cost is calculated and trained until an accurate first conversation record annotation network is generated.

[0161] By repeatedly processing and training multiple such session records, the server continuously improves and optimizes the first session record annotation network, enabling it to more accurately understand and annotate various power business interaction sessions.

[0162] Suppose there is another conversation record: the user said, "Our community often has power outages, what's going on?", and the customer service replied, "It may be a line failure, we are investigating." The first example conversation semantic understanding vector sequence is [v81, v82, v83, ..., v110].

[0163] The server still operates according to the established steps, gradually generates accurate conversation topic classification prediction data, calculates and uses various network learning costs for iterative training, and finally generates the first conversation record annotation network that can effectively handle this type of conversation.

[0164] In actual power business processing, the server continuously receives new session records and uses the trained first session record annotation network to quickly and accurately annotate and process them, providing strong support for the improvement and optimization of power services.

[0165] For example, when the new conversation is "I want to know the specific rules of the tiered electricity price," the server can quickly use the first conversation record annotation network to accurately determine its topic category and provide an accurate basis for subsequent service responses.

[0166] For another example, for a conversation such as "How long will it take to repair a power failure after reporting it?", the server can also accurately mark and process it through the first conversation record marking network, and provide the user with a satisfactory answer in a timely manner.

[0167] Thus, an efficient and accurate first session record annotation network can be generated to better serve the interaction and processing of power business.

[0168] In a possible implementation, step B120 includes: Step C110: Load the first sample conversation semantic understanding vector sequence into the basic neural network layer and the basic topic classification function layer to generate conversation topic classification prediction data, and calculate the second network learning cost based on the conversation topic classification prediction data and the topic summary distribution.

[0169] Step C120: arbitrarily select X sample conversation semantic understanding vectors from the first sample conversation semantic understanding vector sequence, and fuse the X sample conversation semantic understanding vectors according to the conversation logic order to generate a first conversation semantic understanding vector to be learned, where X is a positive integer.

[0170] Step C130: Load the first to-be-learned conversation semantic understanding vector into the first neural network layer and the first basic semantic extraction function layer in the at least one basic semantic extraction function layer to generate a first conversation semantic estimation vector.

[0171] Step C140: Calculate a third network learning cost based on the first conversation semantics estimation vector and the X+1th sample conversation semantics understanding vector in the first sample conversation semantics understanding vector sequence.

[0172] Step C150: randomly determine a set number of first semantic feature nodes from the first sample conversation semantic understanding vector sequence, perform masking processing on them, and generate a second conversation semantic understanding vector to be learned.

[0173] Step C160: Load the second to-be-learned conversation semantic understanding vector into the second neural network layer and the second basic semantic extraction function layer in the at least one basic semantic extraction function layer to generate a second conversation semantic estimation vector.

[0174] Step C170: calculating a fourth network learning cost based on the second conversation semantic estimation vector and the actual feature vector of the first semantic feature node.

[0175] Step C180: training the basic neural network layer and the basic topic classification function layer based on the second network learning cost, the third network learning cost and the fourth network learning cost to generate a temporary conversation record annotation network in the first round of training.

[0176] Step C190: when the temporary session record annotated network meets the network convergence condition, obtain the first session record annotated network.

[0177] In this embodiment, for example, there is a conversation record like this: the user said, "My family's electricity consumption has increased sharply recently. Is there a power leakage?" The customer service responded, "We will arrange personnel to come to your home for inspection. Please check whether all electrical appliances are turned off first." After preliminary processing, the first example conversation semantic understanding vector sequence of this conversation is [v1, v2, v3, ..., v60].

[0178] The server first loads this first example conversation semantic understanding vector sequence into the basic neural network layer and basic topic classification function layer of the basic conversation record annotation network. After calculation and analysis, the conversation topic classification prediction data is generated. Assume that the generated prediction result is [0.6, 0.3, 0.1], which may belong to the abnormal power consumption category, leakage detection category, and other categories respectively.

[0179] Then, it is known that the real topic summary distribution of this session is [1, 0, 0], which clearly indicates that it belongs to the abnormal power consumption category. The server calculates the learning cost of the second network based on the session topic classification prediction data [0.6, 0.3, 0.1] and the topic summary distribution [1, 0, 0].

[0180] Next, the server randomly selects 10 vectors (assuming they are v1-v10) from the first example conversation semantic understanding vector sequence, fuses these 10 vectors according to the logical order of the conversation, and generates the first conversation semantic understanding vector to be learned.

[0181] The first conversation semantic understanding vector to be learned is loaded into the first neural network layer and the first basic semantic extraction function layer, and after processing, a first conversation semantic estimation vector is generated.

[0182] Based on this first conversation semantics estimation vector and the 11th vector (ie, v11) in the first example conversation semantics understanding vector sequence, the server calculates the third network learning cost.

[0183] Afterwards, the server randomly determines 5 semantic feature nodes (assumed to be v20, v30, v40, v50, and v60) from the first example conversation semantic understanding vector sequence and performs masking processing to generate the second conversation semantic understanding vector to be learned.

[0184] The second conversation semantics understanding vector to be learned is loaded into the second neural network layer and the second basic semantics extraction function layer to generate a second conversation semantics estimation vector.

[0185] Based on the second session semantic estimation vector and the actual feature vectors of the five masked semantic feature nodes (v20, v30, v40, v50, v60), the server calculates the fourth network learning cost.

[0186] Based on the second network learning cost, the third network learning cost and the fourth network learning cost, the server trains the basic neural network layer and the basic topic classification function layer to generate a temporary session record annotation network in the first round of training.

[0187] During the training process, the server continuously monitors the performance indicators of the temporary conversation record annotation network. If the gap between the output results of the temporary conversation record annotation network and the actual topic induction distribution gradually narrows after multiple iterations and meets the preset network convergence conditions (for example, the error is less than a certain threshold), then the server obtains the final first conversation record annotation network.

[0188] For example, in another conversation record, the user said, "I want to change the way I pay my electricity bill. How do I do that?" The customer service replied, "You can do it through the mobile app or at the business hall." The first example conversation semantic understanding vector sequence is [v70, v71, v72, ..., v100].

[0189] The server also processes according to the above steps. Load the vector sequence into the base layer to generate prediction data and calculate the learning cost of the second network. Select 10 vectors to fuse to generate the first vector to be learned and calculate the learning cost of the third network. Randomly shield 5 nodes to generate the second vector to be learned and calculate the learning cost of the fourth network. Then, based on these three costs, train and generate a temporary session record annotation network until the convergence conditions are met to obtain the first session record annotation network.

[0190] By repeatedly processing and training a plurality of such session records, the server continuously optimizes and improves the first session record annotation network, so that it can more accurately annotate and analyze various power business interaction sessions.

[0191] Suppose there is another conversation record: the user says, "How long will it take for someone to come and handle the power repair?" The customer service replies, "They will usually arrive within 24 hours." The first example conversation semantic understanding vector sequence is [v110, v111, v112, ..., v150].

[0192] The server still operates according to the same process and finally generates a first session record annotation network that can accurately handle such sessions.

[0193] In actual power business processing, the server uses the continuously optimized first session record annotation network to quickly and accurately understand and annotate various user sessions, thereby improving service efficiency and quality.

[0194] For example, when a new conversation is "I have questions about my electricity bill," the server can quickly use the trained network to accurately determine its topic and provide strong support for subsequent processing.

[0195] For another example, for a conversation such as "How to check the progress of power failure repair?", the server can also accurately mark and respond through the first conversation record marking network.

[0196] In a possible implementation, step B120 may further include: Step D110: arbitrarily select X sample conversation semantic understanding vectors from the first sample conversation semantic understanding vector sequence, and fuse the X sample conversation semantic understanding vectors according to the conversation logic order to generate a first conversation semantic understanding vector to be learned, where X is a positive integer.

[0197] Step D120: Load the first to-be-learned conversation semantic understanding vector into the first basic semantic extraction function layer among the basic neural network layer and the at least one basic semantic extraction function layer to generate a first conversation semantic estimation vector; and calculate the third network learning cost based on the first conversation semantic estimation vector and the X+1th sample conversation semantic understanding vector in the first sample conversation semantic understanding vector sequence.

[0198] Step D130: randomly determine a set number of first semantic feature nodes from the first sample conversation semantic understanding vector sequence, perform shielding processing on them, and generate a second conversation semantic understanding vector to be learned; load the second conversation semantic understanding vector to be learned into the second basic semantic extraction function layer in the basic neural network layer and the at least one basic semantic extraction function layer to generate a second conversation semantic estimation vector.

[0199] Step D140, calculate the fourth network learning cost based on the second conversation semantic estimation vector and the actual feature vector of the first semantic feature node, and train the basic neural network layer, the basic topic classification function layer, the first basic semantic extraction function layer, and the second basic semantic extraction function layer based on the third network learning cost and the fourth network learning cost to generate a first neural network layer, a first topic classification function layer, the first semantic extraction function layer, and the second semantic extraction function layer.

[0200] Step D150: Load the first sample conversation semantic understanding vector sequence into the first neural network layer and the first topic classification function layer to generate conversation topic classification prediction data, and calculate the second network learning cost based on the conversation topic classification prediction data and the topic summary distribution.

[0201] Step D160: arbitrarily select X sample conversation semantic understanding vectors from the first sample conversation semantic understanding vector sequence, and fuse the X sample conversation semantic understanding vectors according to the conversation logic order to generate a third conversation semantic understanding vector to be learned, and load the third conversation semantic understanding vector to be learned into the first neural network layer and the first semantic extraction function layer to generate a third conversation semantic estimation vector.

[0202] Step D170: Calculate a fifth network learning cost based on the third conversation semantics estimation vector and the X+1th sample conversation semantics understanding vector in the first sample conversation semantics understanding vector sequence.

[0203] Step D180: randomly determine a set number of second semantic feature nodes from the first sample conversation semantic understanding vector sequence, perform shielding processing on them, generate a fourth conversation semantic understanding vector to be learned, load the fourth conversation semantic understanding vector to be learned into the first neural network layer and the second semantic extraction function layer to generate a fourth conversation semantic estimation vector, and calculate the sixth network learning cost based on the fourth conversation semantic estimation vector and the actual feature vector of the second semantic feature node.

[0204] Step D190, training the first neural network layer and the first topic classification function layer based on the second network learning cost, the fifth network learning cost and the sixth network learning cost, generating a temporary conversation record annotation network in the first round of training, and obtaining the first conversation record annotation network when the temporary conversation record annotation network meets the network convergence conditions.

[0205] In this embodiment, assume that one of the conversation records is: the user says "The reading of my newly installed electricity meter seems to be inaccurate, can you help check it?", and the customer service replies "OK, we will arrange for professionals to come to check it as soon as possible." After preliminary processing, the first example conversation semantic understanding vector sequence of this conversation is [v1, v2, v3, ..., v80].

[0206] First, the server randomly selects 8 sample conversation semantic understanding vectors (assumed to be v1-v8) from the first sample conversation semantic understanding vector sequence, and fuses these 8 vectors according to the conversation logic order to generate the first conversation semantic understanding vector to be learned.

[0207] Then, the first conversation semantic understanding vector to be learned is loaded into the basic neural network layer and the first basic semantic extraction function layer, and after processing, a first conversation semantic estimation vector is generated.

[0208] Based on this first conversation semantics estimation vector and the 9th vector (ie, v9) in the first example conversation semantics understanding vector sequence, the server calculates the third network learning cost.

[0209] Next, the server randomly determines three first semantic feature nodes (assuming they are v20, v30, and v40) from the first example conversation semantic understanding vector sequence and performs masking processing to generate a second conversation semantic understanding vector to be learned.

[0210] The second conversation semantics understanding vector to be learned is loaded into the basic neural network layer and the second basic semantics extraction function layer to generate a second conversation semantics estimation vector.

[0211] Based on the second session semantic estimation vector and the actual feature vectors of the three masked first semantic feature nodes (v20, v30, v40), the server calculates the fourth network learning cost.

[0212] Based on the third network learning cost and the fourth network learning cost, the server trains the basic neural network layer, the basic topic classification functional layer, the first basic semantic extraction functional layer, and the second basic semantic extraction functional layer to generate the first neural network layer, the first topic classification functional layer, the first semantic extraction functional layer, and the second semantic extraction functional layer.

[0213] Afterwards, the first example conversation semantic understanding vector sequence is loaded into the first neural network layer and the first topic classification function layer to generate conversation topic classification prediction data.

[0214] Based on the generated conversation topic classification prediction data and the known topic summary distribution (for example, [1, 0, 0] represents meter calibration related categories), the server calculates the learning cost of the second network.

[0215] Randomly select 8 sample conversation semantic understanding vectors (assuming they are v10-v17) from the first sample conversation semantic understanding vector sequence, and fuse these 8 vectors according to the conversation logic order to generate the third conversation semantic understanding vector to be learned.

[0216] The third conversation semantic understanding vector to be learned is loaded into the first neural network layer and the first semantic extraction function layer to generate a third conversation semantic estimation vector.

[0217] The fifth network learning cost is calculated based on the third conversation semantics estimation vector and the 18th vector (ie, v18) in the first example conversation semantics understanding vector sequence.

[0218] Three second semantic feature nodes (assuming they are v50, v60, and v70) are randomly determined from the first example conversation semantic understanding vector sequence for masking to generate a fourth conversation semantic understanding vector to be learned.

[0219] The fourth conversation semantic understanding vector to be learned is loaded into the first neural network layer and the second semantic extraction function layer to generate a fourth conversation semantic estimation vector.

[0220] The sixth network learning cost is calculated based on the fourth session semantic estimation vector and the actual feature vectors of the three masked second semantic feature nodes (v50, v60, v70).

[0221] Based on the second network learning cost, the fifth network learning cost and the sixth network learning cost, the first neural network layer and the first topic classification function layer are trained to generate a temporary conversation record annotation network in the first round of training.

[0222] During the training process, the server continuously monitors the performance of the temporary conversation record annotation network. If the error between its output result and the true topic induction distribution gradually decreases and meets the preset network convergence condition (for example, the error is less than a certain threshold), the server obtains the final first conversation record annotation network.

[0223] For example, in another conversation record, the user said, "I want to suspend my electricity use. What procedures are required?" The customer service replied, "You need to provide your ID card, electricity contract, and written application." The first example conversation semantic understanding vector sequence is [v90, v91, v92, ..., v120].

[0224] The server follows the same steps. First, select 8 vectors to fuse to generate the first vector to be learned, and calculate the third network learning cost; randomly mask to generate the second vector to be learned, and calculate the fourth network learning cost; train to generate a new network layer. Then load and generate the conversation topic classification prediction data, calculate the second network learning cost; select 8 vectors to fuse again to generate the third vector to be learned, and calculate the fifth network learning cost; randomly mask to generate the fourth vector to be learned, and calculate the sixth network learning cost. Finally, train based on these costs to generate a temporary conversation record annotation network until the convergence conditions are met, and the first conversation record annotation network is obtained.

[0225] By repeatedly and meticulously processing and training multiple such conversation records, the server continuously optimizes and improves the first conversation record annotation network, enabling it to more accurately understand and annotate various types of power business interactive conversations, providing strong support for the intelligent and efficient development of power services.

[0226] For example, when the new conversation is "I would like to consult the specific regulations of time-of-use electricity prices," the server can quickly and accurately use the trained first conversation record annotation network to determine its topic and provide accurate guidance for subsequent services.

[0227] For another example, for a conversation like "My electric meter seems to be broken, what should I do?", the server can also rely on the first session record annotation network to perform accurate annotation and processing, and provide the user with an effective solution in a timely manner.

[0228] Figure 2 The electric power business interactive conversation processing system 100 based on a large language model shown includes: a processor 1001 and a memory 1003. The processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the electric power business interactive conversation processing system 100 based on a large language model may also include a transceiver 1004, which may be used for data interaction between the server and other servers, such as data transmission and / or data reception. It should be noted that the transceiver 1004 is not limited to one in actual scheduling, and the structure of the electric power business interactive conversation processing system 100 based on a large language model does not constitute a limitation on the embodiments of the present application.

[0229] Processor 1001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 1001 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

[0230] The bus 1002 may include a path to transmit information between the above components. The bus 1002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. The bus 1002 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 2 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0231] The memory 1003 may be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compressed optical disk, laser disk, optical disk, digital versatile disk, Blu-ray disk, etc.), magnetic disk storage media, other magnetic storage devices, or any other medium that can be used to carry or store program code and can be read by a computer, without limitation herein.

[0232] The memory 1003 is used to store program codes for executing the embodiments of the present application, and the execution is controlled by the processor 1001. The processor 1001 is used to execute the program codes stored in the memory 1003 to implement the steps shown in the above method embodiments.

[0233] An embodiment of the present application provides a computer-readable storage medium having program code stored thereon. When the program code is executed by a processor, the steps and corresponding contents of the aforementioned method embodiment can be implemented.

[0234] It should be understood that, although each operation step is characterized by directed lines in the flow chart of the embodiment of the present application, the implementation order of these steps is not limited to the order covered by the directed lines. Unless there is clear explanation in this article, in some implementation scenarios of the embodiment of the present application, the implementation steps in each flow chart can be performed in other orders based on demand. In addition, some or all of the steps in each flow chart may include multiple sub-steps or multiple stages according to actual implementation scenarios, and some or all of these sub-steps or stages may be performed in the same stage, and each sub-step or stage in these sub-steps or stages may also be performed in different stages respectively. Under different scenarios in the execution stage, the execution order of these sub-steps or stages can be flexibly configured based on demand, and the embodiment of the present application does not limit this.

[0235] The above is only an optional implementation method for some implementation scenarios of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present application, other similar implementation methods based on the technical ideas of the present application are also within the protection scope of the embodiments of the present application.

Claims

1. A method for processing electric power business interactive conversations based on a large language model, characterized in that: The method comprises: Obtaining a conversation semantic understanding vector sequence of a target power business interaction conversation record; Performing semantic enhancement processing on the conversation semantic understanding vector sequence to generate a conversation semantic enhancement vector sequence; The conversation semantics enhancement vector sequence is loaded into a first conversation record annotation network to generate conversation record annotation data of the target power business interaction conversation record, wherein the first conversation record annotation network is a network parameter learning generation based on conversation semantics extraction instructions and topic induction instructions, the network loading data of the first conversation record annotation network is a first example learning conversation record sequence, the first example learning conversation record sequence includes a first example conversation semantics understanding vector sequence and a topic induction distribution of the first example conversation semantics understanding vector sequence, the topic induction distribution represents the conversation record annotation data of the first example conversation semantics understanding vector sequence, and the first example conversation semantics understanding vector sequence is a conversation semantics understanding vector that has been semantically enhanced.

2. The method for processing electric power business interactive conversations based on a large language model according to claim 1, characterized in that: The step of performing semantic enhancement processing on the conversation semantic understanding vector sequence to generate a conversation semantic enhancement vector sequence comprises: The conversation semantic understanding vector sequence is loaded into a semantic reinforcement processing model to generate the conversation semantic reinforcement vector sequence, wherein the semantic reinforcement processing model performs self-learning generation based on the second sample conversation semantic understanding vector sequence, the sample conversation semantic reinforcement vector sequence, and the first semantic vector, wherein the first semantic vector is generated by fusion of the second sample conversation semantic understanding vector sequence and the sample conversation semantic reinforcement vector sequence; Alternatively, the conversation semantics understanding vector sequence is masked to generate the conversation semantics enhancement vector sequence.

3. The method for processing electric power business interactive conversations based on a large language model according to claim 2, characterized in that: The method further comprises: Acquire a second sample learning conversation record sequence and a basic semantic reinforcement processing model, wherein the second sample learning conversation record sequence includes the second sample conversation semantic understanding vector sequence, the sample conversation semantic reinforcement vector sequence, and the first semantic vector; The second sample learning conversation record sequence is loaded into the basic semantic reinforcement processing model for semantic reinforcement to generate a second semantic vector, and the second semantic vector is generated by blending simulation based on the second sample conversation semantic understanding vector sequence and the sample conversation semantic reinforcement vector sequence; Calculating a first network learning cost according to the second semantic vector and the first semantic vector; The basic semantic enhancement processing model is trained based on the first network learning cost to generate a semantic enhancement processing model.

4. The method for processing electric power business interactive conversations based on a large language model according to claim 3, characterized in that: The obtaining of the second sample learning session record sequence comprises: Acquire a first sample learning session record, a sample semantic enhancement session record, and a first electric power business interaction session record, wherein the first electric power business interaction session record is generated by merging the first sample learning session record and the sample semantic enhancement session record; In the non-online mode, semantic extraction is performed on the first sample learning session record to generate the second sample session semantic understanding vector sequence, and semantic extraction is performed on the sample semantic enhancement session record to generate the sample session semantic enhancement vector sequence, and semantic extraction is performed on the first power business interaction session record to generate the first semantic vector.

5. The method for processing electric power business interactive conversations based on a large language model according to claim 4, characterized in that: The obtaining of the first sample learning session record, the sample semantic enhancement session record and the first power business interaction session record comprises: Acquire the first example learning session record, and generate a semantic random perturbation session record, wherein the semantic random perturbation session record is used as the sample semantic reinforcement session record; fuse the semantic random perturbation session record with the first example learning session record to generate the first power business interaction session record; Alternatively, the first sample learning session record is obtained, and the sequence of the first sample learning session record is adjusted to generate a rearranged power business interaction session record of the first sample learning session record, and the rearranged power business interaction session record is used as the sample semantically enhanced session record; the rearranged power business interaction session record is merged with the first sample learning session record to generate the first power business interaction session record; Alternatively, the first sample learning session record is obtained, and the local attention weight content is screened on the first sample learning session record to generate a screened power business interaction session record of the first sample learning session record, and the screened power business interaction session record is used as the sample semantic enhancement session record; the screened power business interaction session record is merged with the first sample learning session record to generate the first power business interaction session record; Alternatively, the first sample learning session record is obtained, and the keywords of the first sample learning session record are replaced to generate a keyword replacement session record of the first sample learning session record, and the keyword replacement session record is used as the sample semantic enhancement session record; the keyword replacement session record is merged with the first sample learning session record to generate the first power business interaction session record.

6. The method for processing electric power business interactive conversations based on a large language model according to any one of claims 1 to 5, characterized in that: The method further comprises: Acquire a first example learning conversation record sequence and a basic conversation record annotation network, wherein the second example learning conversation record sequence includes the first example conversation semantic understanding vector sequence and the topic induction distribution of the first example conversation semantic understanding vector sequence, the basic conversation record annotation network includes a basic neural network layer, a basic topic classification function layer and at least one basic semantic extraction function layer, the topic induction distribution represents the conversation record annotation data of the first example conversation semantic understanding vector sequence, the first example conversation semantic understanding vector sequence is a conversation semantic understanding vector that has been semantically enhanced, the processing data of the at least one basic semantic extraction function layer is used to train the basic neural network layer and the basic topic classification function layer, the processing data of the basic topic classification function layer is the processing data of the basic conversation record annotation network, and the first example conversation semantic understanding vector sequence represents the continuous semantic context vector of the electric power business interaction conversation record; The first sample learning conversation record sequence is loaded into the basic conversation record annotation network, and network parameter learning is performed based on the topic induction instruction and the conversation semantic extraction instruction to generate the first conversation record annotation network.

7. The method for processing electric power business interactive conversations based on a large language model according to claim 6, characterized in that: The step of loading the first sample learning conversation record sequence into the basic conversation record annotation network, and performing network parameter learning based on the topic induction instruction and the conversation semantic extraction instruction to generate the first conversation record annotation network comprises: Loading the first sample conversation semantic understanding vector sequence into the basic neural network layer and the basic topic classification function layer to generate conversation topic classification prediction data; Calculating a second network learning cost based on the conversation topic classification prediction data and the topic summary distribution; Iteratively training the basic neural network layer and the basic subject classification function layer based on the second network learning cost to generate a first neural network layer and a first subject classification function layer; Randomly select X sample conversation semantic understanding vectors from the first sample conversation semantic understanding vector sequence, and fuse the X sample conversation semantic understanding vectors according to the conversation logic order to generate a first conversation semantic understanding vector to be learned, where X is a positive integer; Loading the first to-be-learned conversation semantics understanding vector into the first neural network layer and the first basic semantics extraction function layer of the at least one basic semantics extraction function layer to generate a first conversation semantics estimation vector; Calculate a third network learning cost based on the first conversation semantics estimation vector and the X+1th sample conversation semantics understanding vector in the first sample conversation semantics understanding vector sequence; Based on the third network learning cost, the first neural network layer and the first subject classification function layer are trained to generate a second neural network layer and a second subject classification function layer; Randomly determine a set number of first semantic feature nodes from the first sample conversation semantic understanding vector sequence, perform masking processing on them, and generate a second conversation semantic understanding vector to be learned; Loading the second to-be-learned conversation semantics understanding vector into the second neural network layer and the second basic semantics extraction function layer in the at least one basic semantics extraction function layer to generate a second conversation semantics estimation vector; Calculate a fourth network learning cost based on the second session semantic estimation vector and the actual feature vector of the first semantic feature node; The second neural network layer and the second topic classification function layer are trained based on the fourth network learning cost to generate the first conversation record annotation network.

8. The method for processing electric power business interactive conversations based on a large language model according to claim 6, characterized in that: The step of loading the first sample learning conversation record sequence into the basic conversation record annotation network, and performing network parameter learning based on the topic induction instruction and the conversation semantic extraction instruction to generate the first conversation record annotation network comprises: Loading the first sample conversation semantic understanding vector sequence into the basic neural network layer and the basic topic classification function layer to generate conversation topic classification prediction data; Calculating a second network learning cost based on the conversation topic classification prediction data and the topic summary distribution; Randomly select X sample conversation semantic understanding vectors from the first sample conversation semantic understanding vector sequence, and fuse the X sample conversation semantic understanding vectors according to the conversation logic order to generate a first conversation semantic understanding vector to be learned, where X is a positive integer; Loading the first to-be-learned conversation semantics understanding vector into the first neural network layer and the first basic semantics extraction function layer of the at least one basic semantics extraction function layer to generate a first conversation semantics estimation vector; Calculate a third network learning cost based on the first conversation semantics estimation vector and the X+1th sample conversation semantics understanding vector in the first sample conversation semantics understanding vector sequence; Randomly determine a set number of first semantic feature nodes from the first sample conversation semantic understanding vector sequence, perform masking processing on them, and generate a second conversation semantic understanding vector to be learned; Loading the second to-be-learned conversation semantics understanding vector into the second neural network layer and the second basic semantics extraction function layer in the at least one basic semantics extraction function layer to generate a second conversation semantics estimation vector; Calculate a fourth network learning cost based on the second session semantic estimation vector and the actual feature vector of the first semantic feature node; Based on the second network learning cost, the third network learning cost and the fourth network learning cost, the basic neural network layer and the basic topic classification function layer are trained to generate a temporary conversation record annotation network in the first round of training; When the temporary session record annotated network meets the network convergence condition, the first session record annotated network is obtained.

9. The method for processing electric power business interactive conversations based on a large language model according to claim 6, characterized in that: The step of loading the first sample learning conversation record sequence into the basic conversation record annotation network, and performing network parameter learning based on the topic induction instruction and the conversation semantic extraction instruction to generate the first conversation record annotation network comprises: Randomly select X sample conversation semantic understanding vectors from the first sample conversation semantic understanding vector sequence, and fuse the X sample conversation semantic understanding vectors according to the conversation logic order to generate a first conversation semantic understanding vector to be learned, where X is a positive integer; Loading the first to-be-learned conversation semantics understanding vector into the first basic semantics extraction function layer of the basic neural network layer and the at least one basic semantics extraction function layer to generate a first conversation semantics estimation vector; Calculate a third network learning cost based on the first conversation semantics estimation vector and the X+1th sample conversation semantics understanding vector in the first sample conversation semantics understanding vector sequence; Randomly determine a set number of first semantic feature nodes from the first sample conversation semantic understanding vector sequence, perform masking processing on them, and generate a second conversation semantic understanding vector to be learned; Loading the second to-be-learned conversation semantics understanding vector into the second basic semantics extraction function layer in the basic neural network layer and the at least one basic semantics extraction function layer to generate a second conversation semantics estimation vector; Calculate a fourth network learning cost based on the second session semantic estimation vector and the actual feature vector of the first semantic feature node; Based on the third network learning cost and the fourth network learning cost, the basic neural network layer, the basic topic classification function layer, the first basic semantic extraction function layer, and the second basic semantic extraction function layer are trained to generate a first neural network layer, a first topic classification function layer, the first semantic extraction function layer, and the second semantic extraction function layer; Loading the first sample conversation semantic understanding vector sequence into the first neural network layer and the first topic classification function layer to generate conversation topic classification prediction data; Calculating a second network learning cost based on the conversation topic classification prediction data and the topic summary distribution; Randomly selecting X sample conversation semantic understanding vectors from the first sample conversation semantic understanding vector sequence, and fusing the X sample conversation semantic understanding vectors according to the conversation logic order to generate a third conversation semantic understanding vector to be learned; Loading the third to-be-learned conversation semantics understanding vector into the first neural network layer and the first semantic extraction function layer to generate a third conversation semantics estimation vector; Calculate a fifth network learning cost based on the third conversation semantics estimation vector and the X+1th sample conversation semantics understanding vector in the first sample conversation semantics understanding vector sequence; Randomly determine a set number of second semantic feature nodes from the first sample conversation semantic understanding vector sequence, perform masking processing on them, and generate a fourth conversation semantic understanding vector to be learned; Loading the fourth to-be-learned conversation semantics understanding vector into the first neural network layer and the second semantic extraction function layer to generate a fourth conversation semantics estimation vector; Calculating a sixth network learning cost based on the fourth session semantic estimation vector and the actual feature vector of the second semantic feature node; Based on the second network learning cost, the fifth network learning cost and the sixth network learning cost, the first neural network layer and the first topic classification function layer are trained to generate a temporary conversation record annotation network in the first round of training; When the temporary session record annotated network meets the network convergence condition, the first session record annotated network is obtained.

10. A power business interactive conversation processing system based on a large language model, characterized in that: It includes a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by the processor, the method for processing electric power business interactive conversations based on a large language model as described in any one of claims 1 to 9 is implemented.

Images