Power Business Interaction Session Processing Method and System Based on Large Language Model

Abstract

The present application provides a method and system for processing power service interaction sessions based on large language models. First, a sequence of session semantic understanding vectors of a target power service interaction session record is obtained. Then, semantic enhancement processing is performed on it to obtain a sequence of session semantic enhancement vectors. Finally, the enhanced vector sequence is loaded into a first session record annotation network, which is generated by learning network parameters according to session semantic extraction instructions and topic induction instructions, and its loaded data is a first sample learning session record sequence, which includes a first sample session semantic understanding vector sequence after enhancement processing and its topic induction distribution. In this way, session record annotation data of the target power service interaction session record can be generated, so as to realize the accurate analysis and processing of power service interaction sessions, improve the quality and efficiency of power services, and better meet the needs of users in power services.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more particularly, to a method and system for processing power business interaction sessions based on large language models. Background Art

[0002] With the rapid development and intelligent transformation of the power industry, the types and complexity of power business have been continuously increasing, and the interaction sessions between users and power service providers have become increasingly frequent and diverse. In power business, accurately understanding and processing user interaction sessions is crucial for providing high-quality services, solving problems, and optimizing business processes.

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

[0004] Some existing natural language processing techniques may not be able to fully extract the deep semantic information in the interaction sessions when dealing with power business interaction sessions, and the understanding of semantics is not accurate and comprehensive enough. This may lead to misjudgment of user needs and affect service quality and business decisions.

[0005] In addition, when annotating and classifying sessions, there is a lack of effective models and methods, making it difficult to accurately extract the themes and key information of the sessions 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 intelligent level of processing power business interaction sessions, it is necessary to develop an innovative method based on large language models to achieve in-depth understanding and accurate annotation of power business interaction sessions. Summary of the Invention

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

[0008] According to the first aspect of the present application, there is provided a method for processing power business interaction sessions based on large language models, the method comprising:

[0009] Obtaining a sequence of conversation semantic understanding vectors of a target power business interaction session record;

[0010] Performing semantic enhancement processing on the sequence of conversation semantic understanding vectors to generate a sequence of conversation semantic enhancement vectors;

[0011] Load the sequence of session semantic enhancement vectors into the first session record annotation network to generate session record annotation data for the target power service interaction session. The first session record annotation network is generated by learning network parameters based on session semantic extraction instructions and topic induction instructions. The network loading data of the first session record annotation network is a first sample learning session record sequence, which includes a first sample session semantic understanding vector sequence and the topic induction distribution of the first sample session semantic understanding vector sequence. The topic induction distribution represents the session record annotation data of the first sample session semantic understanding vector sequence, and the first sample session semantic understanding vector sequence is a session semantic understanding vector processed by semantic enhancement.

[0012] According to a second aspect of the present application, there is provided a power service interaction session processing system based on a large language model. The power service interaction session processing system based on a large language model includes a machine-readable storage medium and a processor. The machine-readable storage medium stores machine-executable instructions. When the processor executes the machine-executable instructions, the power service interaction session processing system based on a large language model implements the foregoing power service interaction session processing method based on a large language model.

[0013] According to a third aspect of the present application, there is provided a computer-readable storage medium storing computer-executable instructions, which when executed, implement the foregoing power service interaction session processing method based on a large language model.

[0014] According to any of the above aspects, the technical effect of the present application is as follows:

[0015] Through the semantic enhancement processing of the session semantic understanding vector sequence in the embodiments of the present application, the semantic information in the session can be more deeply mined, semantic ambiguity can be reduced, and thus the user's needs and intentions can be more accurately understood. The first session record annotation network learns based on session semantic extraction instructions and topic induction instructions, and combines a rich first sample learning session record sequence, which can generate more accurate session record annotation data for the target power service interaction session, helping to accurately classify and analyze the session. Thus, a large number of power service interaction session records can be quickly processed, adapting to different types and complex levels of session scenarios, providing strong support for the optimization and improvement of power services. That is, by accurately understanding and annotating the session, the problems of users in power services can be timely and effectively responded to and solved, improving user satisfaction and promoting the intelligent and efficient development of power services. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] To more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the accompanying drawings required for the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and thus should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other relevant drawings can also be obtained based on these drawings.

[0017] Figure 1 Flow chart showing the method for processing power service interaction sessions based on a large language model provided by an embodiment of the present application.

[0018] Figure 2 Component structure diagram of a power service interaction session processing system based on a large language model for implementing the above-mentioned method for processing power service interaction sessions based on a large language model provided by an embodiment of the present application. Detailed implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present application based on the accompanying drawings in the embodiments of the present application. It should be understood that the accompanying drawings in the present application only serve the purpose of illustration and description, and are not used to limit the protection scope of the present application. Additionally, it should be understood that the schematic drawings are not drawn to actual scale. The flowcharts used in the present application show the operations implemented according to some embodiments of the embodiments of the present application. It should be understood that the operations in the flowchart may not be implemented in sequence, and steps without logical context relationships may be reversed or implemented simultaneously. Moreover, those skilled in the art can add multiple other operations to the flowchart or delete multiple operations from the flowchart under the guidance of the content of the present application.

[0020] Figure 1 Flow chart showing the method and system for processing power service interaction sessions based on a large language model provided by an embodiment of the present application. It should be understood that in other embodiments, the order of some steps in the method for processing power service interaction sessions based on a large language model in this embodiment can be shared according to actual needs, or some of the steps can also be omitted or maintained. The detailed steps of the method for processing power service interaction sessions based on a large language model include:

[0021] Step S110, obtaining a sequence of session semantic understanding vectors of a target power service interaction session record.

[0022] Specifically, the target power service interaction session record refers to the actual interaction conversation record between users and customer service or the system in the power system regarding power services. The session semantic understanding vector sequence is obtained by converting each word or phrase in this target power service interaction session record into a vector of a fixed length (i.e., word embedding), and encoding these vectors through natural language processing techniques (such as recurrent neural networks or long short-term memory networks), thereby obtaining a vector sequence that can represent the entire session semantics.

[0023] In this embodiment, it is assumed that in the power system where the server is located, a user has an interaction session with a customer service staff regarding electricity bill inquiry. The user said, "I want to inquire about this month's electricity bill. How do I operate?" The customer service staff replied, "You can inquire through our official website or mobile APP. Just enter your user account and password."

[0024] The server first performs text preprocessing on this target power service interaction session record, including removing noise, converting to a unified encoding format, etc. Then, using the word embedding method in natural language processing technology, each word is converted into a vector of a fixed length. Next, these word vectors are encoded through models such as recurrent neural networks (RNN) or long short-term memory networks (LSTM) to obtain the session semantic understanding vector sequence of the entire target power service interaction session record.

[0025] For example, after processing, the user's first sentence "I want to inquire about this month's electricity bill. How do I operate?" may be represented as a sequence of 20 vectors [v1, v2, v3,..., v20], and the dimension of each vector is 100. The customer service staff's reply "You can inquire 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], and the dimension of each vector is also 100.

[0026] Step S120: Perform semantic enhancement processing on the session semantic understanding vector sequence to generate a session semantic enhancement vector sequence.

[0027] Specifically, semantic enhancement processing is a method for further processing the session semantic understanding vector sequence, aiming to enhance or highlight the key semantic information in the session and reduce semantic ambiguity. It can be achieved by using a pre-trained semantic enhancement processing model or performing masking processing (such as randomly selecting vectors for masking). The session semantic enhancement vector sequence is the session semantic understanding vector sequence after semantic enhancement processing, and the vectors in it may contain richer semantic information, which helps to more accurately understand the session content.

[0028] In this embodiment, the server can perform semantic enhancement processing in various ways.

[0029] Method 1: Using a semantic enhancement processing model

[0030] Suppose the server has a pre-trained semantic enhancement processing model. This model is generated through self-learning 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.

[0031] For the conversation semantic understanding vector sequence [v1, v2, v3,..., v50] of the obtained target power service interaction conversation record, the server loads it into the semantic enhancement processing model. Inside the model, these vectors may be further analyzed and processed. For example, through the Attention Mechanism to focus on important semantic parts, or through a multi-layer fully connected neural network to extract higher-level semantic features. After processing, a conversation semantic enhancement vector sequence [v'1, v'2, v'3,..., v'50] is generated.

[0032] Method 2: Performing masking processing

[0033] The server can also directly perform masking processing on the conversation semantic understanding vector sequence to generate a conversation semantic enhancement vector sequence. For example, randomly select some vectors in the sequence for masking, or mask them according to a certain rule (such as masking one vector every 5 vectors). The masked vectors can be replaced with zero vectors or other preset special vectors. After the masking processing, a conversation semantic enhancement vector sequence is obtained.

[0034] Step S130: Load the conversation semantic enhancement vector sequence into the first conversation record annotation network to generate the conversation record annotation data of the target power service interaction conversation record, where the first conversation record annotation network is generated by learning network parameters based on conversation semantic extraction instructions and topic induction instructions, the network loading data of the first conversation record annotation network is the first sample learning conversation record sequence, the first 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 topic induction distribution represents the conversation record annotation data of the first sample conversation semantic understanding vector sequence, and the first sample conversation semantic understanding vector sequence is the conversation semantic understanding vector after semantic enhancement processing.

[0035] Specifically, the first session record annotation network is a neural network model that learns network parameters based on discourse semantic extraction instructions and topic induction instructions. Its function is to classify and annotate the discourse semantic vector sequence after semantic enhancement processing to generate session record annotation data.

[0036] The discourse semantic extraction instruction is an instruction or task that requires the network model to extract key semantic information from the session. This is usually achieved by training the model to identify keywords, phrases, or sentences in the session. The topic induction instruction is an instruction or task that requires the network model to classify the session into specific topics or categories. For example, electricity bill inquiry, fault repair request, new electricity meter application, etc. This is usually achieved by training the model to identify the overall topic or intention of the session.

[0037] The network loading data refers to the data used for training or loaded into the neural network model. In the first session record annotation network, the network loading data is the first example learning session record sequence.

[0038] The first example learning session record sequence is a set of example data for training the first session record annotation network, including the first example discourse semantic understanding vector sequence and the topic induction distribution corresponding to these vector sequences. The first example discourse semantic understanding vector sequence refers to the example data of the discourse semantic understanding vector sequence after semantic enhancement processing, which is used to train the first session record annotation network.

[0039] The topic induction distribution is a vector or probability distribution that represents the possibility of the session record belonging to different topics or categories. For example, [0.9, 0.05, 0.05] indicates that the session has a 90% possibility of belonging to the electricity bill inquiry category.

[0040] That is to say, in this embodiment, the first session record annotation network is generated by learning network parameters based on the discourse semantic extraction instruction and the topic induction instruction. Its network loading data is the first example learning session record sequence, and this first example learning session record sequence includes the first example discourse semantic understanding vector sequence and its topic induction distribution.

[0041] Suppose the first example learning session record sequence contains various conversations about power services, such as electricity bill inquiry, fault repair request, new electricity meter application, etc., and each conversation has a corresponding topic induction distribution. For example, the topic induction distribution of "electricity bill inquiry" may be [1, 0, 0], indicating belonging to the electricity bill inquiry category; the topic induction distribution of "fault repair request" may be [0, 1, 0], indicating belonging to the fault repair request category; the topic induction distribution of "new electricity meter application" may be [0, 0, 1], indicating belonging to the new electricity meter application category.

[0042] The server loads the sequence of session semantic enhancement vectors of the target power service interaction session record into the first session record annotation network. Neurons in the network calculate and process these vectors, and through a multi-layer convolutional neural network (CNN) or Transformer architecture, etc., finally output a predicted topic induction distribution.

[0043] For example, for the above electricity bill query session, after being processed by the first session record annotation network, the output topic induction distribution may be [0.9, 0.05, 0.05], indicating that this session has a 90% probability of belonging to the electricity bill query category, a 5% probability of belonging to the fault repair category, and a 5% probability of belonging to the new electricity meter application category.

[0044] Based on this output topic induction distribution, the server determines the session record annotation data of the target power service interaction session record, that is, it determines that it mainly belongs to the electricity bill query category.

[0045] Suppose in another scenario, the user's conversation with the customer service is about power fault repair. The user says, "The power in my house suddenly went out. I don't know why." The customer service replies, "Please check if the circuit breaker has tripped first. If not, we will arrange for a repairman to come and check as soon as possible."

[0046] The server processes this session according to the above step S110 to obtain the sequence of session semantic understanding vectors. Then in step S120, the sequence of session semantic enhancement vectors is generated through the semantic enhancement processing model or masking processing. Finally, in step S130, it is loaded into the first session record annotation network, and the output topic induction distribution may be [0.02, 0.95, 0.03], thereby determining that this session belongs to the fault repair category.

[0047] Another example is that the user's conversation with the customer service is about applying for a new electricity meter. The user says, "I newly purchased a house and want to apply for a new electricity meter." The customer service replies, "Okay, please prepare the relevant property certificate and ID card, and we will handle it for you."

[0048] The server also goes through the processing of steps S110 to S130, and the finally output topic induction distribution may be [0.01, 0.02, 0.97], determining that this session belongs to the new electricity meter application category.

[0049] With this design, it can effectively obtain the sequence of session semantic understanding vectors of the target power service interaction session record, perform semantic enhancement processing, and use the first session record annotation network to generate accurate session record annotation data, thereby realizing the intelligent analysis and processing of power service interaction sessions.

[0050] Based on the above steps, through the semantic enhancement processing of the conversation semantic understanding vector sequence, the embodiments of the present application can more deeply mine the semantic information in the conversation, reduce semantic ambiguity, and thus more accurately understand the needs and intentions of users. The first conversation record annotation network learns based on the conversation semantic extraction instruction and the topic induction instruction, and combines a rich first sample to learn the conversation record sequence, and can generate more accurate conversation record annotation data for the target power service interaction conversation, which helps to accurately classify and analyze the conversation. Thereby, a large number of power service interaction conversation records can be quickly processed, adapting to different types and complexity levels of conversation scenarios, and providing strong support for the optimization and improvement of power services. That is, by accurately understanding and annotating conversations, problems of users in power services can be timely and effectively responded to and solved, user satisfaction can be improved, and the intelligent and efficient development of power services can be promoted.

[0051] In a possible implementation manner, step S120 includes:

[0052] Step S121, loading the conversation semantic understanding vector sequence into a semantic enhancement processing model to generate the conversation semantic enhancement vector sequence, where the semantic enhancement processing model is self-learned based on a second sample conversation semantic understanding vector sequence, a sample conversation semantic enhancement vector sequence, and a first semantic vector, and the first semantic vector is generated by fusing the second sample conversation semantic understanding vector sequence and the sample conversation semantic enhancement vector sequence.

[0053] Alternatively, step S122, performing a masking process on the conversation semantic understanding vector sequence to generate the conversation semantic enhancement vector sequence.

[0054] In this embodiment, in the power system, the server undertakes the task of processing a large number of power service interaction conversation records.

[0055] Suppose the server receives a target power service interaction conversation record, such as a discussion between a user and a customer service about abnormal power consumption. The user says: "I think my power consumption has suddenly increased a lot this month. Is there something wrong with the electricity meter?" The customer service replies: "Hello, we will arrange staff to check. It may be that some electrical appliances are leaking electricity or your electricity usage habits have changed."

[0056] The server first performs text preprocessing and word embedding operations on this conversation according to the conventional processing flow, and encodes it through models such as a recurrent neural network or a long short-term memory network to obtain a conversation semantic understanding vector sequence, assumed to be [v1, v2, v3,..., v40].

[0057] Method 1: Using a semantic enhancement processing model

[0058] The server has a pre-trained semantic enhancement processing model. This semantic enhancement processing model is generated through self-learning 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.

[0059] Suppose there are numerous second sample conversation semantic understanding vector sequences stored in the server's database, such as:

[0060] - Conversation about electricity bill disputes: [v101, v102, v103,..., v150]

[0061] - Conversation about power outage notices: [v201, v202, v203,..., v250]

[0062] There are also corresponding sample conversation semantic enhancement vector sequences, such as:

[0063] - Enhanced electricity bill dispute conversation: [v151, v152, v153,..., v200]

[0064] - Enhanced power outage notice conversation: [v251, v252, v253,..., v300]

[0065] Fuse these second sample conversation semantic understanding vector sequences and sample conversation semantic enhancement vector sequences to generate the first semantic vectors. For example, fuse them through methods such as weighted average or concatenation.

[0066] For the conversation semantic understanding vector sequence [v1, v2, v3,..., v40] of the currently obtained target power business interaction conversation record, the server loads it into the semantic enhancement processing model. The model may analyze the importance of each vector through an attention mechanism inside.

[0067] For example, in this discussion about abnormal power consumption, vectors corresponding to keywords such as "sudden increase in power consumption" and "problem with the electricity meter" may be given higher weights. Then, use a multi-layer fully connected neural network to extract deeper semantic features.

[0068] After processing, a conversation semantic enhancement vector sequence is generated, assumed to be [v'1, v'2, v'3,..., v'40]. These enhanced vectors may have changed values in certain dimensions or added new semantic information to better reflect the core semantics of the conversation.

[0069] Method 2: Perform masking processing

[0070] The server can also choose to directly perform masking processing on the conversation semantic understanding vector sequence to generate the conversation semantic enhancement vector sequence.

[0071] Suppose the shielding rule set by the server is to shield one vector every three vectors. For the sequence of session semantic understanding vectors [v1, v2, v3,..., v40], vectors such as v4, v7, v10, etc. will be shielded.

[0072] The shielded vectors are replaced with zero vectors. After such processing, a sequence of session semantic enhanced vectors is obtained, such as [v1, v2, v3, 0, v5, v6, v7, 0,...].

[0073] Through this kind of shielding processing, the model may pay more attention to the key information carried by the unshielded vectors in subsequent learning and analysis, thereby improving the ability to capture important semantics.

[0074] Another example is that another segment of the target power business interaction session record is about power equipment maintenance. The user said: "That transformer seems to be malfunctioning and has been making noise." The customer service replied: "We will send someone to check and repair it as soon as possible."

[0075] The server also first obtains the sequence of session semantic understanding vectors, assumed to be [v51, v52, v53,..., v70].

[0076] When using the semantic enhancement processing model, the model may analyze and enhance the current vector sequence based on the previously learned similar maintenance-related example vector sequences and enhanced vector sequences, highlighting key information such as "transformer failure" and "noise".

[0077] When performing the shielding processing, if the rule is to shield one vector every two vectors, then vectors such as v53, v55, v57, etc. will be shielded to generate the corresponding sequence of session semantic enhanced vectors.

[0078] Through the above different ways of semantic enhancement processing, the server can provide more representative and effective input data for the subsequent session record annotation network, thereby improving the analysis and annotation accuracy of power business interaction sessions.

[0079] Suppose there is another session about the inquiry of the progress of a 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 about one month left. It depends on the construction progress specifically."

[0080] After the server obtains its sequence of session semantic understanding vectors [v81, v82, v83,..., v100], if it uses the semantic enhancement processing model, the model may strengthen the semantic vectors related to time and construction according to the past example data of similar construction progress inquiries.

[0081] If masking processing is adopted, for example, masking one vector every four vectors, a new sequence of session semantic enhancement vectors is generated.

[0082] In this way, when processing various types of power business interaction sessions, the server can flexibly apply these two semantic enhancement processing methods to adapt to different session characteristics and requirements, laying a foundation for accurately annotating session records.

[0083] Continue to assume that there is also a session about power complaints. The user said excitedly: "Your service is too bad. I've been waiting for a long time and no one has come to solve the problem!" The customer service quickly apologized and said that it would be processed as soon as possible.

[0084] After the server obtains the session semantic understanding vector sequence, through the semantic enhancement processing model, it can focus on strengthening the user's dissatisfaction and the core issues of the complaint.

[0085] Or through masking processing, selectively ignore some less critical information and highlight the important semantic parts.

[0086] In short, regardless of the type of power business interaction 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.

[0087] Suppose there is a more complex session that involves the detailed discussion of a power contract. The user and the customer service had a detailed exchange on some terms.

[0088] After the server obtains the session semantic understanding vector sequence, using the semantic enhancement processing model, it can comprehensively consider a large number of contract-related sample data accumulated before and extract more accurate semantic features.

[0089] And through masking processing, the data can be simplified and the key information can be made more prominent.

[0090] In the actual operation of the power system, a large number of various session records are generated every day. By continuously applying these semantic enhancement processing methods, the server can gradually improve its ability to understand and analyze sessions, providing strong support for the optimization and improvement of power business.

[0091] For example, for a session about consulting power preferential policies, the server can accurately strengthen the relevant semantics and provide more valuable information for subsequent judgment of the session theme and category.

[0092] Another example is that when processing a session about emergency handling of power failures, through effective semantic enhancement, the server can quickly identify key problems and arrange corresponding handling measures in a timely manner.

[0093] Over time, the sample data accumulated by the server becomes increasingly rich, the semantic enhancement processing model is continuously optimized, and the shielding processing rules are also more reasonable, making the processing of power business interaction sessions more efficient and accurate.

[0094] Suppose that within a certain period, what users mainly consult about is the issue of power price adjustment. Through the semantic enhancement processing of these sessions, the server can better summarize and generalize the concerns and needs of users, providing a reference basis for the power department to formulate relevant policies.

[0095] Suppose there is a discussion session about the access of new energy to the power grid. The server can also extract key information through the above semantic enhancement processing method, providing valuable suggestions for the planning and development of the power system.

[0096] Whether it is the daily consultation of ordinary users or the communication of large-scale electricity consumption needs of enterprise users, the server can ensure the smooth progress of power business interaction through precise semantic enhancement processing, improving service quality and user satisfaction.

[0097] Suppose during a peak electricity consumption period, a large number of users have questions and suggestions about electricity consumption restrictions and allocation. Through fast and accurate semantic enhancement processing, the server can promptly sort out and summarize this information, providing strong data support for power dispatching decisions.

[0098] For another example, when the power system is undergoing technological upgrading and transformation, users inquire about the possible impacts and changes. Through effective semantic enhancement processing, the server can clearly understand the concerns and expectations of users, thereby better communicating and explaining with users.

[0099] With this design, the semantic enhancement processing method can be flexibly applied to handle various complex power business interaction session situations, providing a solid technical guarantee for the stable operation and high-quality service of the power system.

[0100] In a possible implementation manner, the method further includes:

[0101] Step A110, obtaining a second sample learning session record sequence and a basic semantic enhancement processing model, where the second sample learning session record sequence includes the second sample conversation semantic understanding vector sequence, the sample conversation semantic enhancement vector sequence, and the first semantic vector.

[0102] Step A120, loading the second sample learning session record sequence into the basic semantic enhancement processing model to generate a second semantic vector through semantic enhancement. The second semantic vector is generated by fusing and simulating according to the second sample conversation semantic understanding vector sequence and the sample conversation semantic enhancement vector sequence.

[0103] Step A130, calculate the first network learning cost according to the second semantic vector and the first semantic vector.

[0104] Step A140, train the basic semantic enhancement processing model based on the first network learning cost to generate a semantic enhancement processing model.

[0105] First, the server obtains a second sample learning session record sequence, assuming it includes multiple different types of power service session records. For example:

[0106] - A session about electricity bill overdue notice, whose second sample conversation semantic understanding vector sequence is [v1001, v1002, v1003,..., v1050].

[0107] - A session about the arrangement of power equipment upgrade, whose sample conversation semantic enhancement vector sequence is [v2001, v2002, v2003,..., v2050].

[0108] Meanwhile, through the fusion operation of these two sequences, a first semantic vector is generated, for example, obtained through a specific weighted fusion method.

[0109] The server has an initial basic semantic enhancement processing model. Load the obtained second sample learning session record sequence into this basic semantic enhancement processing model.

[0110] Inside the basic semantic enhancement processing model, according to the preset algorithms and parameters, it conducts a blending simulation process on the input second sample conversation semantic understanding vector sequence and the sample conversation semantic enhancement vector sequence. For example, by performing interactive calculations, feature fusion, etc. on the vector elements in the two sequences, a second semantic vector is generated.

[0111] Assume that for the session about electricity bill overdue notice mentioned above, the generated second semantic vector is [v1101, v1102, v1103,..., v1150]; for the session about the arrangement of power equipment upgrade, the generated second semantic vector is [v2101, v2102, v2103,..., v2150].

[0112] Next, the server calculates the first network learning cost according to the generated second semantic vector and the first semantic vector fused previously. The calculation method may be to compare the differences between these two semantic vectors in each dimension, and use cost functions such as mean square error and cross entropy to quantify this difference.

[0113] For example, for the two vectors of the electricity bill overdue notice session, calculate the sum of the squares of the differences of their corresponding elements, and divide by the vector length for averaging.

[0114] Based on the calculated first network learning cost, the server trains the basic semantic enhancement processing model. The weight parameters in the model are adjusted so that when the model processes new example learning session records, the generated second semantic vector can be closer to the first semantic vector, that is, more in line with the expected semantic representation.

[0115] In another scenario, assume that the second example learning session record sequence contains a session about emergency repair of power failures, and its second example session semantic understanding vector sequence is [v3001, v3002, v3003,..., v3050], and the example session semantic enhancement vector sequence is [v4001, v4002, v4003,..., v4050].

[0116] After the fusion simulation processing of the basic semantic enhancement processing model, the generated second semantic vector is [v3101, v3102, v3103,..., v3150].

[0117] Then, the first network learning cost is calculated by comparing with the fused first semantic vector, and the model is further trained and optimized based on this cost.

[0118] In this way, by continuously obtaining new second example learning session record sequences and repeating the above process, the basic semantic enhancement processing model of the server is gradually trained into a mature model that can accurately perform semantic enhancement processing, providing more effective support for subsequent processing of various power business interaction sessions.

[0119] For another example, a session record about consulting on opening an account for new power users is obtained, and it is processed and trained according to the same process to continuously improve the performance and accuracy of the model.

[0120] In actual power business operations, continuously performing such operations enables the model to adapt to the continuously changing and rich power business session scenarios, providing strong guarantees for the efficient operation and high-quality service of the power system.

[0121] In a possible implementation manner, step A110 includes:

[0122] Step A111, obtaining the first example learning session record, the example semantic enhancement session record, and the first power business interaction session record, where the first power business interaction session record is generated by fusing the first example learning session record and the example semantic enhancement session record.

[0123] Step A112, in the non-online mode, perform semantic extraction on the first sample learning conversation record to generate the second sample conversation semantic understanding vector sequence, perform semantic extraction on the sample semantic reinforcement conversation record to generate the sample conversation semantic reinforcement vector sequence, and perform semantic extraction on the first power service interaction conversation record to generate the first semantic vector.

[0124] In this embodiment, first, the server obtains the first sample learning conversation record. For example, a conversation record about a user applying for an increase in power capacity. The user says, "I've added a lot of new appliances at home, and the existing power capacity is not enough. I want to apply for an increase."

[0125] At the same time, the server obtains the sample semantic reinforcement conversation record. For example, the sample semantic reinforcement conversation record generated by performing keyword replacement on the first sample learning conversation record. Replace "power capacity" with "power consumption", becoming "I've added a lot of new appliances at home, and the existing power consumption is not enough. I want to apply for an increase."

[0126] Then, fuse the first sample learning conversation record and the sample semantic reinforcement conversation record to generate the first power service interaction conversation record.

[0127] In the non-online mode, the server performs semantic extraction on the first sample learning conversation record. For example, use a word vector model and a neural network to process the above conversation record for applying for an increase in capacity, and convert it into a series of digital vectors, thereby generating the second sample conversation semantic understanding vector sequence [v1, v2, v3,..., vn].

[0128] Next, perform the same semantic extraction operation on the sample semantic reinforcement conversation record to generate the sample conversation semantic reinforcement vector sequence [v'1, v'2, v'3,..., v'm].

[0129] Finally, perform semantic extraction on the fused first power service interaction conversation record to generate the first semantic vector [v''1, v''2, v''3,..., v''k].

[0130] For another example, in another scenario, the first sample learning conversation record is about a user consulting the tiered electricity price policy, "I'm not very clear about how the tiered electricity price is divided here." The sample semantic reinforcement conversation record is obtained by adjusting its order, "How is the tiered electricity price divided here? I'm not very clear."

[0131] After the server fuses these two conversation records, perform semantic extraction respectively in the non-online mode to generate the corresponding second sample conversation semantic understanding vector sequence, sample conversation semantic reinforcement vector sequence, and first semantic vector.

[0132] For another example, the first sample learning conversation record is that the user feedbacks that the electricity meter reading is abnormal, "The electricity meter reading this month seems incorrect. It is much higher than before." The sample semantic enhancement conversation record is obtained by generating semantic random perturbations. For example, "The electricity meter reading this month seems not quite right. It is much, much higher than usual." After fusion, semantic extraction is performed to obtain the required vector sequence and semantic vector.

[0133] By continuously obtaining and processing various different first sample learning conversation records and sample semantic enhancement conversation records, the data samples can be enriched, providing strong support for subsequent model training and optimization.

[0134] In a possible implementation manner, step A111 includes:

[0135] Obtain the first sample learning conversation record and generate a semantic random perturbation conversation record, and the semantic random perturbation conversation record serves as the sample semantic enhancement conversation record. Fuse the semantic random perturbation conversation record with the first sample learning conversation record to generate the first power service interaction conversation record.

[0136] Or, obtain the first sample learning conversation record and perform sequential adjustment on the first sample learning conversation record to generate a rearranged power service interaction conversation record of the first sample learning conversation record, and the rearranged power service interaction conversation record serves as the sample semantic enhancement conversation record. Fuse the rearranged power service interaction conversation record with the first sample learning conversation record to generate the first power service interaction conversation record.

[0137] Or, obtain the first sample learning conversation record and perform local attention weight content screening on the first sample learning conversation record to generate a screened power service interaction conversation record of the first sample learning conversation record, and the screened power service interaction conversation record serves as the sample semantic enhancement conversation record. Fuse the screened power service interaction conversation record with the first sample learning conversation record to generate the first power service interaction conversation record.

[0138] Or, obtain the first sample learning conversation record and perform keyword replacement on the first sample learning conversation record to generate a keyword replacement conversation record of the first sample learning conversation record, and the keyword replacement conversation record serves as the sample semantic enhancement conversation record. Fuse the keyword replacement conversation record with the first sample learning conversation record to generate the first power service interaction conversation record.

[0139] In the power system, the server needs to obtain and process various sample learning conversation records for subsequent analysis and model training.

[0140] The first case:

[0141] The server first obtains a first sample learning session record. For example, the user says, "Why is the electricity bill so high this month in my home? Is there something wrong with the electricity meter?" Then, the server generates a semantically randomly perturbed session record by randomly changing some words or sentence structures. For example, "Why is the electricity bill so high this month in my home? Could it be that there is something wrong with the electricity meter?" This semantically randomly perturbed session record serves as the sample semantic reinforcement session record. Finally, the semantically randomly perturbed session record is fused with the original first sample learning session record to generate the first power service interaction session record.

[0142] The second case:

[0143] The server obtains the first sample learning session record as: "I want to install a new electricity meter. What materials do I need to prepare?" Then, this session record is reordered to generate a reordered power service interaction session record. For example, "What materials do I need to prepare? I want to install a new electricity meter." This reordered power service interaction session record serves as the sample semantic reinforcement session record and is fused with the original first sample learning session record to obtain the first power service interaction session record.

[0144] The third case:

[0145] The server obtains the first sample learning session record as: "The community has been experiencing frequent power outages recently, which has affected normal life." Then, local attention weight content screening is performed on it. For example, the key part "the community has been experiencing frequent power outages" is screened out to generate a screened power service interaction session record: "The community has been experiencing frequent power outages." This screened record serves as the sample semantic reinforcement session record and is fused with the first sample learning session record to form the first power service interaction session record.

[0146] The fourth case:

[0147] The server obtains the first sample learning session record as: "The process of power repair is too complicated. Can it be simplified?" After that, the keywords in it are replaced. For example, "power repair" is replaced with "power maintenance" to generate a keyword replacement session record: "The process of power maintenance is too complicated. Can it be simplified?" This record serves as the sample semantic reinforcement session record and is fused with the original first sample learning session record to obtain the first power service interaction session record.

[0148] Through the above different methods, the server obtains rich and diverse sample data, providing more comprehensive and effective data support for subsequent power service session analysis and processing.

[0149] In a possible implementation manner, the method further includes:

[0150] Step B110: Obtain the first sample learning session record sequence and the basic session record annotation network. Among them, the second sample learning session 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 session 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 session 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 undergone semantic enhancement processing. 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 session record annotation network. The first sample conversation semantic understanding vector sequence represents the continuous semantic context vector of the power service interaction session record.

[0151] Step B120: Load the first sample learning session record sequence into the basic session record annotation network, and perform network parameter learning based on the topic induction instruction and the conversation semantic extraction instruction to generate the first session record annotation network.

[0152] In this embodiment, first, the server obtains the first sample learning session record sequence. For example, there is a series of session records:

[0153] - Session 1: The user said, "I want to query the electricity bill details for last month. How can I do that?" The customer service replied, "You can query by entering the electricity meter number through the mobile APP or the online business hall." After processing, the first sample conversation semantic understanding vector sequence of this session may be [v101, v102, v103,..., v150], and its topic induction distribution is [1, 0, 0], indicating that it belongs to the electricity bill query category.

[0154] - Session 2: The user said, "The utility pole near my home seems to have fallen. It's very dangerous." The customer service responded, "We will send someone to check and handle it immediately." The first sample conversation semantic understanding vector sequence of this session may be [v201, v202, v203,..., v250], and the topic induction distribution is [0, 1, 0], indicating that it belongs to the power equipment failure category.

[0155] At the same time, the server also obtains the basic session 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 session. The multiple basic semantic extraction function layers are used to extract deeper semantic information from the vector.

[0156] Next, the server loads the first sample learning session record sequence into the basic session record annotation network.

[0157] For session one, the server loads its first sample 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. Suppose the prediction data is [0.8, 0.1, 0.1], and then the server calculates the second network learning cost based on this prediction data and the known topic induction distribution [1, 0, 0]. If the prediction data has a large gap with the true topic induction distribution, for example, in this case, the prediction probability of the electricity bill query category is 0.8 instead of 1, then the learning cost will be relatively high.

[0158] Based on the calculated second network learning cost, the server iteratively trains the basic neural network layer and the basic topic classification function layer, adjusts the network parameters, and generates a more accurate first neural network layer and first topic classification function layer.

[0159] Then, the server randomly selects 5 vectors (assumed to be v101 - v105) from the first sample session semantic understanding vector sequence of session one, and fuses these 5 vectors according to the session logic order to generate the first to-be-learned session semantic understanding vector. This first to-be-learned session semantic understanding vector is loaded into the first neural network layer and the first basic semantic extraction function layer to generate the first session semantic estimation vector. The third network learning cost is calculated based on this first session semantic estimation vector and the 6th vector (i.e., v106) in the first sample session semantic understanding vector sequence.

[0160] Based on the third network learning cost, the first neural network layer and the first topic classification function layer are trained again to generate the second neural network layer and the second topic classification function layer.

[0161] After that, the server randomly determines 3 semantic feature nodes (assumed to be v110, v120, v130) in the first sample session semantic understanding vector sequence of session one for masking processing to generate the second to-be-learned session semantic understanding vector. This second to-be-learned session semantic understanding vector is loaded into the second neural network layer and the second basic semantic extraction function layer to generate the 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 3 masked semantic feature nodes (v110, v120, v130).

[0162] Train the second neural network layer and the second topic classification function layer based on the fourth network learning cost, gradually optimize the performance of the network, and finally generate the first session record annotation network.

[0163] For Session 2, the server processes and trains according to the same process. Load its first sample conversation semantic understanding vector sequence [v201, v202, v203,..., v250] into the network, generate session topic classification prediction data, calculate the second network learning cost and perform corresponding training. Then select and fuse vectors, calculate the third network learning cost, perform masking processing, calculate the fourth network learning cost, and continuously optimize the network parameters to generate the first session record annotation network that can accurately annotate power equipment failure conversations.

[0164] Another example is another session record. The user says, "I want to handle the time-of-use electricity price business. What procedures are required?" The customer service replies, "You need to provide your ID card, property ownership certificate, and recent electricity bills." Assume that the first sample conversation semantic understanding vector sequence of this session is [v301, v302, v303,..., v350], and the topic induction distribution is [0, 0, 1], indicating that it belongs to the electricity price business handling category.

[0165] The server also loads it into the basic session record annotation network and processes and trains according to the above steps, continuously adjusting the network parameters so that the network can accurately identify and annotate the topics of such sessions.

[0166] By repeatedly loading, calculating the learning cost, and training multiple such first sample learning session record sequences, the server continuously optimizes the parameters of the basic session record annotation network and generates the first session record annotation network that can accurately understand and annotate various power business interaction sessions.

[0167] 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.

[0168] For example, when there is a new user session "I have a question about this month's electricity bill. Can you help me explain it?" The server can quickly process it into a conversation semantic understanding vector sequence and load it into the trained first session record annotation network, output the accurate topic induction distribution, determine that it belongs to the electricity bill query category, and thus provide clear guidance for subsequent services.

[0169] Another example is that the new session is "Our factory wants to carry out power expansion. How do we apply?" The server can also accurately determine that it belongs to the power business application category through the first session record annotation network and take corresponding handling measures.

[0170] Over time, the server can continuously obtain new sequences of first-sample learning session records, update and optimize the first session record annotation network, so that it can always adapt to the development and changes of power services, and provide users with more high-quality and efficient services.

[0171] Suppose that during the peak summer electricity consumption period, the sessions of a large number of users involve issues such as increased electricity bills and electricity consumption restrictions. Through the first session record annotation network, the server can quickly classify and annotate these sessions, timely summarize the focus of users' attention and needs, and provide data support for the power department to formulate reasonable coping strategies.

[0172] Another example is that when the power department launches new preferential policies, the number of user consultation sessions increases. The server uses the first session record annotation network to accurately identify the themes of these sessions and provide users with accurate and detailed policy interpretations and handling guides.

[0173] In summary, by obtaining the first-sample 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 achieve intelligent and efficient processing of power service interaction sessions, and improve the quality of power services and user satisfaction.

[0174] In a possible implementation manner, step B120 includes:

[0175] Step B121, loading the first-sample session semantic understanding vector sequence into the basic neural network layer and the basic theme classification function layer to generate session theme classification prediction data.

[0176] Step B122, calculating a second network learning cost based on the session theme classification prediction data and the theme induction distribution, and performing iterative training on the basic neural network layer and the basic theme classification function layer based on the second network learning cost to generate a first neural network layer and a first theme classification function layer.

[0177] Step B123, arbitrarily selecting X sample session semantic understanding vectors from the first-sample session semantic understanding vector sequence, and fusing the X sample session semantic understanding vectors according to the session logic order to generate a first to-be-learned session semantic understanding vector, where X is a positive integer.

[0178] Step B124, loading the first to-be-learned session semantic understanding vector into the first neural network layer and the first basic semantic extraction function layer among the at least one basic semantic extraction function layer to generate a first session semantic estimation vector.

[0179] Step B125, calculate a third network learning cost based on the first conversation semantic estimation vector and the (X + 1)-th example conversation semantic understanding vector in the first example conversation semantic understanding vector sequence.

[0180] Step B126, train 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.

[0181] Step B127, randomly determine a set number of first semantic feature nodes from the first example conversation semantic understanding vector sequence for masking processing to generate a second conversation semantic understanding vector to be learned.

[0182] Step B128, load the second conversation semantic understanding vector to be learned 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.

[0183] Step B129, calculate a 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 second neural network layer and the second topic classification function layer based on the fourth network learning cost to generate the first conversation record annotation network.

[0184] First, the server obtains a series of first example learning conversation record sequences. For example, there is such a conversation record: The user says, "The electricity bill of my family is too high this month. Is there something wrong with the electricity meter reading?" The customer service replies, "We will arrange personnel to check. Please don't worry for now." After processing, the first example conversation semantic understanding vector sequence of this conversation is [v1, v2, v3,..., v50].

[0185] The server loads this first example conversation semantic understanding vector sequence into the basic neural network layer and the 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 according to the processing results. Suppose the generated prediction data is [0.7, 0.2, 0.1], indicating that there is a 70% possibility of being related to electricity bill issues, a 20% possibility of being other categories, and a 10% possibility of being another category.

[0186] Then, the server compares this conversation topic classification prediction data with the known topic induction distribution (such as [1, 0, 0], indicating that it is determined to be the electricity bill problem category) to calculate the second network learning cost. If the prediction data has a large gap from the true topic induction distribution, then the learning cost will be higher.

[0187] Based on the calculated second network learning cost, the server iteratively trains the basic neural network layer and the basic topic classification function layer. By adjusting the weights and parameters in the network, a more optimized first neural network layer and first topic classification function layer are generated.

[0188] Next, the server randomly selects 5 vectors (assumed to be v1 - v5) from this first sample conversation semantic understanding vector sequence, fuses these 5 vectors in the logical order of the conversation to generate a first to - be - learned conversation semantic understanding vector. Then, this first to - be - learned conversation semantic understanding vector 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.

[0189] After that, the server calculates the third network learning cost based on this first conversation semantic estimation vector and the 6th vector (i.e., v6) in the first sample conversation semantic understanding vector sequence.

[0190] According to the third network learning cost, the first neural network layer and the first topic classification function layer are trained again for further optimization, generating a second neural network layer and a second topic classification function layer.

[0191] Subsequently, the server randomly determines 3 semantic feature nodes (assumed to be v10, v20, v30) from the first sample conversation semantic understanding vector sequence for masking processing to generate a second to - be - learned conversation semantic understanding vector.

[0192] This second to - be - learned conversation semantic understanding vector is loaded into the second neural network layer and the second basic semantic extraction function layer to generate a second conversation semantic estimation vector.

[0193] Next, based on the second conversation semantic estimation vector and the actual feature vectors of the 3 masked semantic feature nodes (v10, v20, v30), the fourth network learning cost is calculated.

[0194] Finally, based on the fourth network learning cost, the second neural network layer and the second topic classification function layer are trained, continuously optimizing and adjusting the network parameters, and ultimately generating a first conversation record annotation network that can accurately label and understand such conversations.

[0195] Another example of a conversation record: The user says "I want to apply for a new electricity meter. What materials do I need to prepare?", and the customer service replies "You need to bring your ID card, property ownership certificate, and application form." Its first sample conversation semantic understanding vector sequence is [v51, v52, v53,..., v80].

[0196] The server also processes according to the above steps. First, it loads into the basic neural network layer and the basic topic classification function layer to generate session topic classification prediction data, then calculates the second network learning cost and conducts corresponding iterative training. Then, it selects 5 vectors for fusion to generate the first session semantic understanding vector to be learned, calculates the third network learning cost and trains. Then, it randomly masks 3 semantic feature nodes to generate the second session semantic understanding vector to be learned, calculates the fourth network learning cost and trains until an accurate first session record annotation network is generated.

[0197] 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.

[0198] Suppose there is another session record: The user says, "There are frequent power outages in our community. What's going on?" The customer service replies, "It may be a line fault and we are checking." Its first sample session semantic understanding vector sequence is [v81, v82, v83,..., v110].

[0199] The server still operates according to the established steps, gradually generates accurate session topic classification prediction data, calculates and uses various network learning costs for iterative training, and finally generates a first session record annotation network that can effectively process such sessions.

[0200] In actual power business processing, the server continuously receives new session records and uses the trained first session record annotation network for fast and accurate annotation and processing, providing strong support for the improvement and optimization of power services.

[0201] For example, when the new session is "I want to know the specific rules of the stepped electricity price.", the server can quickly use the first session record annotation network to accurately judge its topic category and provide an accurate basis for subsequent service responses.

[0202] Another example, for a session like "How long will it take to repair after reporting a power failure?", the server can also accurately annotate and process it through the first session record annotation network and promptly provide a satisfactory answer to the user.

[0203] Thus, an efficient and accurate first session record annotation network can be generated to better serve the interaction and processing of power business.

[0204] In a possible implementation manner, step B120 includes:

[0205] Step C110: Load the first example 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 induction distribution.

[0206] Step C120: Arbitrarily select X example conversation semantic understanding vectors from the first example conversation semantic understanding vector sequence, and fuse the X example conversation semantic understanding vectors according to the conversation logic order to generate the first to-be-learned conversation semantic understanding vector, where X is a positive integer.

[0207] 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 among the at least one basic semantic extraction function layer to generate the first conversation semantic estimation vector.

[0208] Step C140: Calculate the third network learning cost based on the first conversation semantic estimation vector and the (X + 1)-th example conversation semantic understanding vector in the first example conversation semantic understanding vector sequence.

[0209] Step C150: Randomly determine a set number of first semantic feature nodes from the first example conversation semantic understanding vector sequence for masking processing to generate the second to-be-learned conversation semantic understanding vector.

[0210] 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 among the at least one basic semantic extraction function layer to generate the second conversation semantic estimation vector.

[0211] Step C170: 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.

[0212] Step C180: Train 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 stage.

[0213] Step C190: Obtain the first conversation record annotation network when the temporary conversation record annotation network meets the network convergence condition.

[0214] In this embodiment, for example, there is such a conversation record: The user said, "The electricity consumption in my home has increased sharply recently. Is there a leakage?" The customer service replied, "We will arrange for someone to come to your home for inspection. Please check first if all your electrical appliances are turned off." After preliminary processing, the semantic understanding vector sequence of the first example conversation in this conversation is [v1, v2, v3,..., v60].

[0215] The server first loads this semantic understanding vector sequence of the first example conversation into the basic neural network layer and the basic theme classification function layer of the basic conversation record annotation network. Through calculation and analysis, conversation theme classification prediction data is generated. Suppose the generated prediction results are [0.6, 0.3, 0.1], which respectively represent the possible belonging to the electricity consumption anomaly category, the leakage detection category, and other categories.

[0216] Then, it is known that the true theme induction distribution of this conversation is [1, 0, 0], indicating that it clearly belongs to the electricity consumption anomaly category. The server calculates the second network learning cost based on the conversation theme classification prediction data [0.6, 0.3, 0.1] and the theme induction distribution [1, 0, 0].

[0217] Next, the server randomly selects any 10 vectors (assumed to be v1 - v10) from the semantic understanding vector sequence of the first example conversation, and fuses these 10 vectors according to the logical order of the conversation to generate the first to-be-learned conversation semantic understanding vector.

[0218] Load this first to-be-learned conversation semantic understanding vector into the first neural network layer and the first basic semantic extraction function layer, and after processing, generate the first conversation semantic estimation vector.

[0219] Based on this first conversation semantic estimation vector and the 11th vector (i.e., v11) in the semantic understanding vector sequence of the first example conversation, the server calculates the third network learning cost.

[0220] After that, the server randomly determines 5 semantic feature nodes (assumed to be v20, v30, v40, v50, v60) from the semantic understanding vector sequence of the first example conversation for masking processing, generating the second to-be-learned conversation semantic understanding vector.

[0221] Load this second to-be-learned conversation semantic understanding vector into the second neural network layer and the second basic semantic extraction function layer to generate the second conversation semantic estimation vector.

[0222] Based on the second conversation semantic estimation vector and the actual feature vectors of the 5 masked semantic feature nodes (v20, v30, v40, v50, v60), the server calculates the fourth network learning cost.

[0223] 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 the training phase.

[0224] During the training process, the server continuously monitors the performance metrics of the temporary session record annotation network. If, after multiple iterations of the temporary session record annotation network, the gap between its output result and the true topic induction distribution gradually narrows and meets the preset network convergence condition (for example, the error is less than a certain threshold), then the server obtains the final first session record annotation network.

[0225] Another example of a session record: The user says, "I want to change the way I pay my electricity bill. How do I do it?" The customer service replies, "You can do it through the mobile APP or go to the business hall." Its first sample conversation semantic understanding vector sequence is [v70, v71, v72,..., v100].

[0226] The server processes it in the same steps as above. Load the vector sequence into the basic layer to generate prediction data and calculate the second network learning cost. Select 10 vectors to fuse and generate the first vector to be learned, and calculate the third network learning cost. Randomly mask 5 nodes to generate the second vector to be learned, and calculate the fourth network learning cost. Then, based on these three costs, train to generate a temporary session record annotation network until the convergence condition is met to obtain the first session record annotation network.

[0227] By repeatedly processing and training multiple such session records, the server continuously optimizes and improves the first session record annotation network, enabling it to more accurately annotate and analyze various power business interaction sessions.

[0228] Suppose there is another session record: The user says, "How long will it take for someone to come and handle the electricity repair after reporting?" The customer service answers, "Generally, it will arrive within 24 hours." Its first sample conversation semantic understanding vector sequence is [v110, v111, v112,..., v150].

[0229] The server still operates according to the same process and finally generates the first session record annotation network that can accurately handle such sessions.

[0230] In actual power business processing, the server uses the continuously optimized first session record annotation network to quickly and accurately understand and annotate various sessions of users, improving service efficiency and quality.

[0231] For example, when the new session is "I have a question about my electricity bill," the server can quickly use the trained network to accurately judge its topic and provide strong support for subsequent processing.

[0232] For another example, for a conversation such as "How can I query the progress of power failure repair?", the server can also perform accurate annotation and response through the first conversation record annotation network.

[0233] In a possible implementation, step B120 may further include:

[0234] Step D110: Arbitrarily select X example conversation semantic understanding vectors from the first example conversation semantic understanding vector sequence, and fuse the X example conversation semantic understanding vectors according to the conversation logic order to generate a first to-be-learned conversation semantic understanding vector, where X is a positive integer.

[0235] 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 a third network learning cost based on the first conversation semantic estimation vector and the (X + 1)-th example conversation semantic understanding vector in the first example conversation semantic understanding vector sequence.

[0236] Step D130: Randomly determine a set number of first semantic feature nodes from the first example conversation semantic understanding vector sequence for masking processing to generate a second to-be-learned conversation semantic understanding vector, and load the second to-be-learned conversation semantic understanding vector into the second basic semantic extraction function layer among the basic neural network layer and the at least one basic semantic extraction function layer to generate a second conversation semantic estimation vector.

[0237] Step D140: Calculate a 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.

[0238] Step D150: Load the first example 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 a second network learning cost based on the conversation topic classification prediction data and the topic induction distribution.

[0239] 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.

[0240] Step D170 , calculating 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.

[0241] 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.

[0242] 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 to generate a temporary conversation record annotation network in the first round of training. When the temporary conversation record annotation network meets the network convergence conditions, the first conversation record annotation network is obtained.

[0243] In this example, assume that one of the conversation records is: the user says, "My newly installed electricity meter seems to be inaccurate. Can you help check it?" The customer service representative replies, "Okay, we will arrange for a professional to come and check it as soon as possible." After preliminary processing, the first example of this conversation's semantic understanding vector sequence is [v1, v2, v3, ..., v80].

[0244] First, the server randomly selects 8 sample conversation semantic understanding vectors (assuming they are v1-v8) from this 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.

[0245] 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.

[0246] Based on this first conversation semantics estimation vector and the 9th vector (i.e., v9) in the first example conversation semantics understanding vector sequence, the server calculates the third network learning cost.

[0247] Next, the server randomly determines 3 first semantic feature nodes (assumed to be v20, v30, v40) from the first sample conversation semantic understanding vector sequence for masking processing to generate a second to-be-learned conversation semantic understanding vector.

[0248] Load this second to-be-learned conversation semantic understanding vector into the basic neural network layer and the second basic semantic extraction functional layer to generate a second conversation semantic estimation vector.

[0249] Based on the second conversation semantic estimation vector and the actual feature vectors of the 3 masked first semantic feature nodes (v20, v30, v40), the server calculates the fourth network learning cost.

[0250] 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.

[0251] After that, load the first sample conversation semantic understanding vector sequence into the first neural network layer and the first topic classification functional layer to generate conversation topic classification prediction data.

[0252] Based on the generated conversation topic classification prediction data and the known topic induction distribution (for example, [1, 0, 0] represents the category related to meter calibration), the server calculates the second network learning cost.

[0253] Arbitrarily select 8 more sample conversation semantic understanding vectors (assumed to be v10 - v17) from the first sample conversation semantic understanding vector sequence, and fuse these 8 vectors according to the conversation logic order to generate a third to-be-learned conversation semantic understanding vector.

[0254] Load the third to-be-learned conversation semantic understanding vector into the first neural network layer and the first semantic extraction functional layer to generate a third conversation semantic estimation vector.

[0255] Based on the third conversation semantic estimation vector and the 18th vector (i.e., v18) in the first sample conversation semantic understanding vector sequence, calculate the fifth network learning cost.

[0256] Randomly determine 3 second semantic feature nodes (assumed to be v50, v60, v70) from the first sample conversation semantic understanding vector sequence for masking processing to generate a fourth to-be-learned conversation semantic understanding vector.

[0257] Load the fourth to-be-learned conversation semantic understanding vector into the first neural network layer and the second semantic extraction functional layer to generate a fourth conversation semantic estimation vector.

[0258] Calculate the sixth network learning cost based on the fourth conversation semantic estimation vector and the actual feature vectors of the three masked second semantic feature nodes (v50, v60, v70).

[0259] Train 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 to generate a temporary conversation record annotation network in the first round of the training phase.

[0260] 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 specific threshold), the server obtains the final first conversation record annotation network.

[0261] Another example of a conversation record: The user says, "I want to handle the business of suspending power supply. What procedures are required?" The customer service replies, "You need to provide your ID card, power supply contract, and written application." Its first sample conversation semantic understanding vector sequence is [v90, v91, v92,..., v120].

[0262] The server processes it according to the same steps. First, select 8 vectors to fuse and generate the first vector to be learned, calculate the third network learning cost; randomly mask to generate the second vector to be learned, calculate the fourth network learning cost; train to generate a new network layer. Then load to generate 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, calculate the fifth network learning cost; randomly mask to generate the fourth vector to be learned, calculate the sixth network learning cost. Finally, train based on these costs to generate a temporary conversation record annotation network until the convergence condition is met to obtain the first conversation record annotation network.

[0263] 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 service interaction conversations, providing strong support for the intelligence and efficiency of power services.

[0264] For example, when the new conversation is "I want to consult the specific regulations on time-of-use electricity price.", the server can quickly and accurately use the trained first conversation record annotation network to judge its topic and provide accurate guidance for subsequent services.

[0265] Another example, for a conversation like "My electricity meter seems to be broken. What should I do?", the server can also rely on the first conversation record annotation network for accurate annotation and processing, and promptly provide effective solutions for users.

[0266] Figure 2 The large language model-based power service interaction session processing system 100 shown in the figure includes: a processor 1001 and a memory 1003. Among them, the processor 1001 and the memory 1003 are connected, such as connected through a bus 1002. Optionally, the large language model-based power service interaction session processing system 100 may further include a transceiver 1004, and the transceiver 1004 may be used for data interaction between this server and other servers, such as sending and / or receiving data, etc. It should be noted that in actual scheduling, the transceiver 1004 is not limited to one, and the structure of the large language model-based power service interaction session processing system 100 does not constitute a limitation to the embodiments of the present application.

[0267] The processor 1001 may be a CPU (Central Processing Unit, central processor), a general-purpose processor, a DSP (Digital Signal Processor, data signal processor), an ASIC (Application Specific Integrated Circuit, application-specific integrated circuit), an FPGA (Field Programmable Gate Array, field programmable gate array) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It can implement or execute various exemplary logic blocks, modules and circuits described in combination with the disclosure of the present application. The processor 1001 may also be a combination of computing functions, such as a combination including one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0268] The bus 1002 may include a path for transmitting information between the above components. The bus 1002 may be a PCI (Peripheral Component Interconnect, peripheral component interconnect standard) bus or an EISA (Extended Industry Standard Architecture, extended industry standard architecture) bus, etc. The bus 1002 may be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, Figure 2 only a thick line is used to represent it in the figure, but it does not mean that there is only one bus or one type of bus.

[0269] The memory 1003 can be a ROM (ReadOnlyMemory), or other types of static storage devices that can store static information and instructions, a RAM (RandomAccessMemory), or other types of dynamic storage devices that can store information and instructions. It can also be an EEPROM (ElectricallyErasableProgrammableReadOnlyMemory), a CD-ROM (CompactDiscReadOnlyMemory), or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, 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, which is not limited herein.

[0270] The memory 1003 is used to store the program code for implementing the embodiments of the present application and is controlled by the processor 1001 for execution. The processor 1001 is used to execute the program code stored in the memory 1003 to implement the steps shown in the foregoing method embodiments.

[0271] The embodiments of the present application provide a computer-readable storage medium on which program code is stored. When the program code is executed by a processor, the steps and corresponding contents of the foregoing method embodiments can be implemented.

[0272] It should be understood that although the flowcharts of the embodiments of the present application represent various operation steps through directed connection lines, the execution order of these steps is not limited to the order covered by the directed connection lines. Unless clearly stated herein, in some implementation scenarios of the embodiments of the present application, the implementation steps in each flowchart can be executed in other orders based on requirements. In addition, some or all of the steps in each flowchart may include multiple sub-steps or multiple stages according to the actual implementation scenarios. Some or all of these sub-steps or stages can be executed in the same stage, and each sub-step or stage among these sub-steps or stages can also be executed in different stages respectively. In scenarios with different execution stages, the execution order of these sub-steps or stages can be flexibly configured based on requirements, and the embodiments of the present application do not limit this.

[0273] The above are only optional implementation manners of some implementation scenarios of the present application. It should be noted that for those of ordinary skill in the art, without departing from the technical concept of the solution of the present application, adopting other similar implementation means based on the technical idea of the present application also belongs to the protection scope of the embodiments of the present application.

Claims

1. A method for processing power service interaction sessions based on large language models, characterized in that, The method comprises: Obtaining a sequence of conversation semantic understanding vectors 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 for the target power business interaction conversation record, wherein the first conversation record annotation network performs network parameter learning and generation based on conversation semantics extraction instructions and topic induction instructions, and the network loaded 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 undergone semantic enhancement processing.

2. The method for processing power service interaction sessions based on large language models according to claim 1, wherein, The performing semantic enhancement processing on the conversation semantic understanding vector sequence to generate a conversation semantic enhancement vector sequence includes: Loading the conversation semantic understanding vector sequence into a semantic reinforcement processing model to generate the conversation semantic reinforcement vector sequence, the semantic reinforcement processing model self-learning and generating the vector sequence based on the second sample conversation semantic understanding vector sequence, the sample conversation semantic reinforcement vector sequence, and the first semantic vector, where the first semantic vector is generated by fusing 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 power service interaction sessions based on a large language model according to claim 2, wherein 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; Loading the second sample learning conversation record sequence into the basic semantic enhancement processing model for semantic enhancement to generate a second semantic vector, wherein the second semantic vector is generated by blending and simulating the second sample conversation semantic understanding vector sequence and the sample conversation semantic enhancement vector sequence; Calculating a first network learning cost based on 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 power service interaction sessions based on large language models according to claim 3, wherein, The obtaining of the second sample learning session record sequence includes: Obtaining a first sample learning session record, a sample semantic enhancement session record, and a first electric power business interaction session record, where the first electric power business interaction session record is generated by fusing 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 conversation semantic understanding vector sequence, semantic extraction is performed on the sample semantic reinforcement session record to generate the sample conversation semantic reinforcement vector sequence, and semantic extraction is performed on the first power service interaction session record to generate the first semantic vector.

5. The method for processing power service interaction sessions based on large language models according to claim 4, wherein, The obtaining of the first sample learning session record, the sample semantic reinforcement session record, and the first power service interaction session record includes: Obtaining the first sample learning session record and generating a semantic random perturbation session record, where the semantic random perturbation session record serves as the sample semantic reinforcement session record; fusing the semantic random perturbation session record with the first sample learning session record to generate the first power service interaction session record; Alternatively, obtaining the first sample learning session record and performing a sequential adjustment on the first sample learning session record to generate a rearranged power service interaction session record of the first sample learning session record, where the rearranged power service interaction session record serves as the sample semantic reinforcement session record; fusing the rearranged power service interaction session record with the first sample learning session record to generate the first power service interaction session record; Alternatively, obtaining the first sample learning session record and performing a local attention weight content screening on the first sample learning session record to generate a screened power service interaction session record of the first sample learning session record, where the screened power service interaction session record serves as the sample semantic reinforcement session record; fusing the screened power service interaction session record with the first sample learning session record to generate the first power service interaction session record; Alternatively, obtaining the first sample learning session record and performing a keyword replacement on the first sample learning session record to generate a keyword replacement session record of the first sample learning session record, where the keyword replacement session record serves as the sample semantic reinforcement session record; fusing the keyword replacement session record with the first sample learning session record to generate the first power service interaction session record.

6. The method for processing power service interaction sessions based on large language models according to any one of claims 3-5, characterized in that, The method further includes: Obtain the first sample learning session record sequence and the basic session record annotation network. Among them, the second sample learning session 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 session 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 session 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 undergone semantic enhancement processing. 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 session record annotation network. The first sample conversation semantic understanding vector sequence represents the continuous semantic context vector of the power service interaction session record; Load the first sample learning session record sequence into the basic session record annotation network, and perform network parameter learning based on the topic induction instruction and the conversation semantic extraction instruction to generate the first session record annotation network.

7. The method for processing power service interaction sessions based on large language models according to claim 6, wherein, The step of loading the first sample learning session record sequence into the basic session record annotation network and performing network parameter learning based on the topic induction instruction and the conversation semantic extraction instruction to generate the first session record annotation network includes: Load the first sample conversation semantic understanding vector sequence into the basic neural network layer and the basic topic classification function layer to generate session topic classification prediction data; Calculate the second network learning cost based on the session topic classification prediction data and the topic induction distribution; Perform iterative training on the basic neural network layer and the basic topic classification function layer based on the second network learning cost to generate the first neural network layer and the first topic classification function layer; 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 the first to-be-learned conversation semantic understanding vector, where X is a positive integer; Load the first to-be-learned conversation semantic understanding vector into the first basic semantic extraction function layer among the first neural network layer and the at least one basic semantic extraction function layer to generate the first conversation semantic estimation vector; Calculate the third network learning cost based on the first conversation semantic estimation vector and the (X + 1)-th sample conversation semantic understanding vector in the first sample conversation semantic understanding vector sequence; Train the first neural network layer and the first topic classification function layer based on the third network learning cost to generate the second neural network layer and the second topic classification function layer; Randomly determine a set number of first semantic feature nodes from the first sample conversation semantic understanding vector sequence for masking processing to generate the second to-be-learned conversation semantic understanding vector; Loading the second semantic understanding vector of the conversation to be learned into the second neural network layer and the second basic semantic extraction function layer among the at least one basic semantic extraction function layer to generate a second conversation semantic estimation vector; 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; Training the second neural network layer and the second topic classification function layer based on the fourth network learning cost to generate the first conversation record annotation network.

8. The method for processing power service interaction sessions based on large language models according to claim 7, 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 includes: 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 induction distribution; Arbitrarily 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 first semantic understanding vector of the conversation to be learned, where X is a positive integer; Loading the first semantic understanding vector of the conversation to be learned into the first neural network layer and the first basic semantic extraction function layer among the at least one basic semantic extraction function layer to generate a first conversation semantic estimation vector; Calculating a third network learning cost based on the first conversation semantic estimation vector and the (X + 1)-th sample conversation semantic understanding vector in the first sample conversation semantic understanding vector sequence; Randomly determining a set number of first semantic feature nodes from the first sample conversation semantic understanding vector sequence for masking processing to generate a second semantic understanding vector of the conversation to be learned; Loading the second semantic understanding vector of the conversation to be learned into the second neural network layer and the second basic semantic extraction function layer among the at least one basic semantic extraction function layer to generate a second conversation semantic estimation vector; 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; 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 stage; When the temporary conversation record annotation network meets the network convergence condition, obtaining the first conversation record annotation network.

9. The method for processing power service interaction sessions based on large language models according to claim 6, wherein 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 includes: Arbitrarily select X example utterance semantic understanding vectors from the first example utterance semantic understanding vector sequence, and fuse the X example utterance semantic understanding vectors according to the conversation logic order to generate a first to-be-learned utterance semantic understanding vector, where X is a positive integer; Load the first to-be-learned utterance 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 utterance semantic estimation vector; Calculate a third network learning cost based on the first utterance semantic estimation vector and the (X + 1)-th example utterance semantic understanding vector in the first example utterance semantic understanding vector sequence; Randomly determine a set number of first semantic feature nodes from the first example utterance semantic understanding vector sequence for masking processing to generate a second to-be-learned utterance semantic understanding vector; Load the second to-be-learned utterance semantic understanding vector into the second basic semantic extraction function layer among the basic neural network layer and the at least one basic semantic extraction function layer to generate a second utterance semantic estimation vector; Calculate a fourth network learning cost based on the second utterance semantic estimation vector and the actual feature vector of the first semantic feature nodes; 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, a first semantic extraction function layer, and a second semantic extraction function layer; Load the first example utterance semantic understanding vector sequence into the first neural network layer and the first topic classification function layer to generate conversation topic classification prediction data; Calculate a second network learning cost based on the conversation topic classification prediction data and the topic induction distribution; Arbitrarily select X example utterance semantic understanding vectors from the first example utterance semantic understanding vector sequence, and fuse the X example utterance semantic understanding vectors according to the conversation logic order to generate a third to-be-learned utterance semantic understanding vector; Load the third to-be-learned utterance semantic understanding vector into the first neural network layer and the first semantic extraction function layer to generate a third utterance semantic estimation vector; Calculate a fifth network learning cost based on the third utterance semantic estimation vector and the (X + 1)-th example utterance semantic understanding vector in the first example utterance semantic understanding vector sequence; Randomly determine a set number of second semantic feature nodes from the first example utterance semantic understanding vector sequence for masking processing to generate a fourth to-be-learned utterance semantic understanding vector; Load the fourth to-be-learned utterance semantic understanding vector into the first neural network layer and the second semantic extraction function layer to generate a fourth utterance semantic estimation vector; Calculate a sixth network learning cost based on the fourth utterance semantic estimation vector and the actual feature vector of the second semantic feature nodes; Train 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 to generate a temporary session record annotation network in the first round of training phase; When the temporary session record annotation network meets the network convergence condition, obtain the first session record annotation network.

10. A power business interaction session processing system based on a large language model, characterized in that, It includes a processor and a computer-readable storage medium. The computer-readable storage medium stores machine-executable instructions. When the machine-executable instructions are executed by the processor, the method for processing power service interaction sessions based on a large language model according to any one of claims 1-9 is implemented.

Images