Electric power company intelligent outbound method and system based on artificial intelligence

Abstract

The invention provides an intelligent outbound method and system for an electric power company based on artificial intelligence, and relates to the technical field of intelligent outbound, and the method comprises the steps: obtaining user data of the electric power company, constructing a multi-modal user feature representation learning model for a user feature data set, constructing a user-behavior bipartite graph through a graph neural network algorithm, and obtaining a user-behavior bipartite graph; determining a behavior feature representation vector of the user and generating a user portrait vector fused with multi-dimensional information; mapping a user portrait vector to a semantic space of a power field knowledge graph by utilizing a knowledge graph embedding technology, and realizing deep fusion of user features and power business knowledge; in combination with an outbound service scene, an attention mechanism and a strategy gradient algorithm are utilized to autonomously learn and generate personalized outbound strategies oriented to different user groups, and the personalized outbound strategies are subjected to simulation evaluation and dynamic optimization; and converting the personalized outbound strategy into a structured outbound task script, and issuing the structured outbound task script to an intelligent outbound platform.

Description

Technical Field

[0001] The present invention relates to intelligent outbound calling technology, and in particular to an intelligent outbound calling method and system for electric power companies based on artificial intelligence. Background Art

[0002] Power companies are facing increasing customer service demands and massive business processing tasks. Traditional manual customer service and outbound call methods can no longer meet the business needs of modern power companies, and there are problems such as high labor costs, low efficiency, and unstable service quality.

[0003] The prior art has the following defects:

[0004] Poor versatility and difficulty meeting the specific needs of the power industry. Existing intelligent customer service and outbound call systems are mostly designed for general scenarios and are difficult to directly apply to the business scenarios of power companies. The power industry has the characteristics of complex business, strong professionalism, and a wide range of fields involved, which requires specialized knowledge bases and algorithms to support it.

[0005] The degree of personalization is insufficient and the user experience needs to be improved. The outbound call scripts and strategies generated by the existing system are mostly rigid and stereotyped, lacking personalized customization for different user characteristics and needs, resulting in poor user experience and poor outbound call results.

[0006] The semantic understanding ability is limited, and it is difficult to accurately grasp the user's intention. The power field involves a large number of professional terms and complex semantics. The natural language processing ability of the existing system is limited, and it is difficult to accurately understand the intention of the user's inquiry, which affects the service quality. Summary of the invention

[0007] The embodiments of the present invention provide an intelligent outbound calling method and system for a power company based on artificial intelligence, which can solve the problems in the prior art.

[0008] According to a first aspect of the embodiments of the present invention,

[0009] Provides an AI-based intelligent outbound calling method for power companies, including:

[0010] Obtain user data from the power company, the user data including basic information of the user, electricity usage information, payment information, and interactive service records; use data mining technology to perform data cleaning, data integration, data transformation, and data reduction on the user data to generate a standardized user feature data set;

[0011] For the user feature data set, a multimodal user feature representation learning model is constructed. Through the graph neural network algorithm, a user-behavior bipartite graph is constructed to determine the user's behavior feature representation vector and generate a user portrait vector that integrates multi-dimensional information; using the knowledge graph embedding technology, the user portrait vector is mapped to the semantic space of the knowledge graph in the power field to achieve a deep integration of user features and power business knowledge;

[0012] In combination with the outbound call business scenario, the attention mechanism and policy gradient algorithm are used to autonomously learn and generate personalized outbound call strategies for different user groups, and the personalized outbound call strategies are simulated, evaluated and dynamically optimized; the personalized outbound call strategies are converted into structured outbound call task scripts, and sent to the intelligent outbound call platform. The intelligent outbound call platform uses domain ontology and semantic parsing technology to build a spoken dialogue generation model for power business, and automatically generates semantically coherent and logically self-consistent outbound call scripts.

[0013] In an optional embodiment,

[0014] For the user feature data set, a multimodal user feature representation learning model is constructed, and a user-behavior bipartite graph is constructed through a graph neural network algorithm to determine the user's behavior feature representation vector and generate a user portrait vector that integrates multi-dimensional information, including:

[0015] Constructing a user-behavior bipartite graph, wherein the user-behavior bipartite graph includes user nodes and behavior nodes, and there are directed edges between user nodes and behavior nodes; there are undirected edges between behavior nodes, indicating that there is a correlation between different behavior categories; the behavior categories include user payment records, power load curves, and interactive service logs;

[0016] Calculate the weight of the associated edge between the behavior nodes, comprehensively consider the co-occurrence frequency and time sequence pattern of the behavior, and determine the weight of the associated edge based on the co-occurrence frequency of the behavior in the user set, the total number of users, the similarity of the behavior in time, and the balance factor;

[0017] A multi-layer graph attention neural network is designed on the user-behavior bipartite graph. The node hidden state is updated by aggregating neighbor node information, and the importance weights of different neighbor nodes are adaptively calculated through the attention mechanism to learn the high-level semantic representation of the user node and the behavior node. Each layer of the graph attention neural network includes a graph convolution layer and an attention layer. The graph convolution layer propagates node features through graph convolution, and the attention layer calculates the attention weights between nodes by introducing an attention weight vector.

[0018] After obtaining high-level semantic representations of user nodes and behavior nodes through the graph attention neural network learning, a behavior feature fusion layer is designed, and an attention pooling mechanism is adopted to adaptively assign weights to different behavior nodes according to the strength of association between the behavior nodes and the user, and the hidden states of the behavior nodes are weighted and fused to obtain the user's behavior feature representation vector;

[0019] The user behavior feature representation vector and the user attribute feature representation vector are concatenated and fused to construct a multimodal feature representation of the user as a user portrait vector, wherein the user attribute feature representation vector includes static attribute information of the user, including but not limited to age, gender, region, and electricity customer type.

[0020] In an optional embodiment,

[0021] By using knowledge graph embedding technology, the user portrait vector is mapped to the semantic space of the knowledge graph in the power field, realizing the deep integration of user characteristics and power business knowledge, including:

[0022] Construct a time-series dynamic knowledge graph, discretize the continuous time axis into a series of time slices, divide entities, relationships, and triples into corresponding time slices according to the timestamp of the data, construct a static knowledge subgraph on each time slice, and add time-series evolution edges between adjacent time slices to describe the change rules of entities and relationships over time, forming a complete time-series dynamic knowledge graph;

[0023] In each time slice, the TransE model considering time attenuation is used to optimize the time-aware embedding loss function. By minimizing the time-aware embedding loss function in each time slice, the time-related entity embedding representation is obtained. At the same time, the time-smoothing regularization term is introduced to consider the smoothness constraints between adjacent time slices to ensure the continuity of entity and relationship embedding in the time dimension. By jointly optimizing the embedding loss function and time-smoothing constraints of each time slice, the time-series knowledge representation is obtained, which not only retains the static semantic information of each time slice, but also obtains the dynamic evolution law across time slices.

[0024] Given a user's multimodal feature representation at a specific time slice, it is converted into a knowledge embedding space through nonlinear mapping, and the attention weights between the user features and each entity embedding are calculated to indicate the user's attention to different entities.

[0025] Introducing the temporal attention mechanism in the time dimension, calculating the temporal weight of each time slice, dynamically adjusting the knowledge weights of different time slices, comprehensively considering the entity attention weight and the temporal attention weight, enhancing the temporal knowledge representation, and obtaining the user's temporal knowledge enhanced feature representation;

[0026] The time series knowledge enhanced feature representation is spliced ​​with user features to obtain the final dynamic user representation, realizing the deep integration of user features and power business knowledge.

[0027] In an optional embodiment,

[0028] The joint optimization of the embedding loss function and the temporal smoothing constraints of each time slice includes:

[0029]

[0030] Where L represents the loss value at the kth time step, (h, r, t)∈T t represents the true triple set T at the tth time step t A triple in , where h represents the head entity, r represents the relationship, and t represents the tail entity, (h′, r, t′)∈T′ t Represents the negative sampling triple set T′ at the tth time step t A triple in , where h′ represents the head entity of negative sampling, t′ represents the tail entity of negative sampling, y represents the margin parameter used to control the interval between positive and negative triplets, d() represents the distance function, and w k represents the time weight vector of the kth time step, [] + represents the operation of taking the positive part, (h+r)⊙w k represents the element-wise multiplication of the head entity representation under relation r with the temporal weight vector to obtain the time-aware head entity representation, t⊙w k represents the element-wise multiplication of the embedding representation of the tail entity with the temporal weight vector to obtain the time-aware tail entity representation, (h′+r)⊙w k It means to multiply the representation of the negative sampling head entity under the relation r by the time weight vector at the element level to obtain the time-aware negative sampling head entity, t′⊙w k It means to perform element-wise multiplication of the embedding representation of the negative sampling tail entity with the time weight vector to obtain the time-aware negative sampling tail entity;

[0031]

[0032] Among them, R represents the regularization term, K represents the total number of time steps, |E k | represents the entity set E at the kth time step k The size of represents the embedding representation of the i-th entity at the k-th time step, Represents the embedding representation of the i-th relation at the k-th time step.

[0033] In an optional embodiment,

[0034] Combined with the outbound call business scenario, the attention mechanism and policy gradient algorithm are used to autonomously learn and generate personalized outbound call strategies for different user groups, and the personalized outbound call strategies are simulated, evaluated and dynamically optimized, including:

[0035] Define a set of outbound call strategy elements, including outbound call time, frequency, and content. Each strategy element has multiple optional values ​​that constitute its value space. For a given user, its personalized outbound call strategy is defined as a joint probability distribution over a set of strategy elements, which indicates the probability of adopting a specific combination of strategy elements given the user's characteristics.

[0036] Construct a personalized strategy generation model based on the attention mechanism. Take user features as input, dynamically associate user features with strategy elements through the attention mechanism, and use the attention mechanism to calculate the conditional probability distribution of each strategy element under given user features. Use the attention scoring function to measure the correlation between user features and strategy elements. According to different user features, dynamically adjust the importance weight of each strategy element to generate a personalized strategy distribution.

[0037] In the model training phase, the maximum likelihood estimation method is used to minimize the negative log-likelihood loss function. The attention parameters are learned through the stochastic gradient descent optimization algorithm, so that the personalized strategy generation model can generate personalized outbound call strategies that meet the user's preferences based on the user's characteristics.

[0038] The policy gradient algorithm of reinforcement learning is introduced. Through trial and error learning in a simulation environment, each outbound call attempt is defined as a sequential decision-making process interacting with the user. The state is the user characteristics at the current moment and the information of historical outbound call records. The action is the outbound call strategy sampled from the strategy distribution. The reward is the effect evaluation of the outbound call. The outbound call trajectory data is collected by interacting with the environment.

[0039] According to the policy gradient theorem, the parameters of the personalized strategy generation model are updated along the gradient direction to maximize the expected cumulative reward. The state-action value function is approximately estimated using the Monte Carlo method. The estimated value is substituted into the policy gradient formula to obtain the update rule for the parameters of the personalized strategy generation model. The personalized strategy generation model is optimized based on the outbound call trajectory data.

[0040] In an optional embodiment,

[0041] The personalized outbound call strategy is converted into a structured outbound call task script and sent to the intelligent outbound call platform. The intelligent outbound call platform uses domain ontology and semantic parsing technology to build a spoken dialogue generation model for power business, and automatically generates semantically coherent and logically self-consistent outbound call scripts, including:

[0042] Convert the personalized outbound call strategy into a structured task script. The task script adopts a hierarchical structure and contains information about the target user, outbound call time, script template, and business process of the outbound call task. The script template defines the overall structure and slots of the outbound call script, and the business process defines the business process and corresponding API interface that need to be executed during the outbound call process.

[0043] Construct an ontology knowledge base covering electricity consumption, payment, and energy conservation in the power sector. The ontology knowledge base contains the core concepts, attributes, and relationships in the field.

[0044] Perform semantic analysis on outbound call task scripts, identify key entities, attributes, and intentions in the scripts, and map them to corresponding concepts in the ontology knowledge base, associating corresponding attributes and measures;

[0045] A spoken dialogue generation model is constructed using a pre-trained language model. The domain ontology and semantic parsing results are used as input. A template-based generation method is adopted, and the fixed speech fragments in the task script are used as the template skeleton. The language model is used to dynamically fill in the slots in the template to generate complete outbound call speech that is semantically coherent and logically self-consistent.

[0046] According to a second aspect of the embodiments of the present invention,

[0047] Provide an AI-based intelligent outbound call system for power companies, including:

[0048] The first unit is used to obtain user data of the power company, wherein the user data includes basic information of the user, electricity usage information, payment information, and interactive service records; and use data mining technology to perform data cleaning, data integration, data transformation, and data reduction on the user data to generate a standardized user feature data set;

[0049] The second unit is used to construct a multimodal user feature representation learning model for the user feature data set, construct a user-behavior bipartite graph through a graph neural network algorithm, determine the user's behavior feature representation vector and generate a user portrait vector that integrates multi-dimensional information; use knowledge graph embedding technology to map the user portrait vector to the semantic space of the knowledge graph in the power field, and realize the deep integration of user features and power business knowledge;

[0050] The third unit is used to combine the outbound call business scenario, use the attention mechanism and policy gradient algorithm, autonomously learn and generate personalized outbound call strategies for different user groups, and simulate, evaluate and dynamically optimize the personalized outbound call strategies; convert the personalized outbound call strategies into structured outbound call task scripts, and send them to the intelligent outbound call platform. The intelligent outbound call platform uses domain ontology and semantic parsing technology to build a spoken dialogue generation model for power business, and automatically generates semantically coherent and logically self-consistent outbound call scripts.

[0051] According to a third aspect of the embodiments of the present invention,

[0052] An electronic device is provided, comprising:

[0053] processor;

[0054] a memory for storing processor-executable instructions;

[0055] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0056] A fourth aspect of the embodiments of the present invention is:

[0057] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0058] The multimodal user feature representation learning model in this paper can fully mine and integrate the user's feature information in different modes, characterize the user's characteristics from the two perspectives of static attributes and dynamic behaviors, and generate a comprehensive and accurate user portrait vector representation.

[0059] This paper uses knowledge graph embedding technology to map user features to the semantic space of the knowledge graph in the power field, and realizes the deep integration of user features and power business knowledge through the attention mechanism. The fused dynamic user representation not only contains the user's static attributes and dynamic behavior characteristics, but also incorporates the background knowledge in the power field. It has stronger semantic expression ability and interpretability, and provides richer and more accurate user portrait information for the subsequent generation of personalized outbound call strategies.

[0060] This paper proposes a personalized outbound call strategy generation method based on reinforcement learning. Through the attention mechanism and policy gradient algorithm, it autonomously learns and generates personalized outbound call strategies for different user groups, and continuously improves the effect and benefits of outbound call strategies through simulation evaluation and dynamic optimization. This method makes full use of user characteristics and power business knowledge, can adapt to the preferences and feedback of different users, and provides a new idea and implementation solution for the intelligent outbound call system. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A schematic diagram of a flow chart of an intelligent outbound calling method for a power company based on artificial intelligence according to an embodiment of the present invention;

[0062] Figure 2 The figure is a schematic diagram of the structure of an intelligent outbound calling system for a power company based on artificial intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0065] Figure 1 FIG. 1 is a flow chart of an intelligent outbound calling method for a power company based on artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0066] S101. Obtain user data from the power company, the user data including basic information, electricity usage information, payment information, and interactive service records of the user; use data mining technology to perform data cleaning, data integration, data transformation, and data specification on the user data to generate a standardized user feature data set;

[0067] S102. For the user feature data set, a multimodal user feature representation learning model is constructed, and a user-behavior bipartite graph is constructed through a graph neural network algorithm to determine the user's behavior feature representation vector and generate a user portrait vector that integrates multi-dimensional information; using knowledge graph embedding technology, the user portrait vector is mapped to the semantic space of the knowledge graph in the power field to achieve deep integration of user features and power business knowledge;

[0068] S103. Combined with the outbound call business scenario, the attention mechanism and the policy gradient algorithm are used to autonomously learn and generate personalized outbound call strategies for different user groups, and the personalized outbound call strategies are simulated, evaluated and dynamically optimized; the personalized outbound call strategies are converted into structured outbound call task scripts, which are sent to the intelligent outbound call platform. The intelligent outbound call platform uses domain ontology and semantic parsing technology to build a spoken dialogue generation model for the power business, and automatically generates semantically coherent and logically self-consistent outbound call scripts.

[0069] In an optional embodiment,

[0070] For the user feature data set, a multimodal user feature representation learning model is constructed, and a user-behavior bipartite graph is constructed through a graph neural network algorithm to determine the user's behavior feature representation vector and generate a user portrait vector that integrates multi-dimensional information, including:

[0071] Constructing a user-behavior bipartite graph, wherein the user-behavior bipartite graph includes user nodes and behavior nodes, and there are directed edges between user nodes and behavior nodes; there are undirected edges between behavior nodes, indicating that there is a correlation between different behavior categories; the behavior categories include user payment records, power load curves, and interactive service logs;

[0072] Calculate the weight of the associated edge between the behavior nodes, comprehensively consider the co-occurrence frequency and time sequence pattern of the behavior, and determine the weight of the associated edge based on the co-occurrence frequency of the behavior in the user set, the total number of users, the similarity of the behavior in time, and the balance factor;

[0073] A multi-layer graph attention neural network is designed on the user-behavior bipartite graph. The node hidden state is updated by aggregating neighbor node information, and the importance weights of different neighbor nodes are adaptively calculated through the attention mechanism to learn the high-level semantic representation of the user node and the behavior node. Each layer of the graph attention neural network includes a graph convolution layer and an attention layer. The graph convolution layer propagates node features through graph convolution, and the attention layer calculates the attention weights between nodes by introducing an attention weight vector.

[0074] After obtaining high-level semantic representations of user nodes and behavior nodes through the graph attention neural network learning, a behavior feature fusion layer is designed, and an attention pooling mechanism is adopted to adaptively assign weights to different behavior nodes according to the strength of association between the behavior nodes and the user, and the hidden states of the behavior nodes are weighted and fused to obtain the user's behavior feature representation vector;

[0075] The user behavior feature representation vector and the user attribute feature representation vector are concatenated and fused to construct a multimodal feature representation of the user as a user portrait vector, wherein the user attribute feature representation vector.

[0076] For example, for the user feature dataset, this paper proposes a multimodal user feature representation learning model based on graph neural network, which generates a vector representation that comprehensively describes the user portrait by integrating the user's static attribute features and dynamic behavior features. The specific method steps are as follows:

[0077] First, a user-behavior bipartite graph is constructed based on the user feature dataset. The bipartite graph contains two types of nodes: user nodes and behavior nodes. User nodes represent each independent user, and behavior nodes represent different types of user behaviors. There is a directed edge between user nodes and behavior nodes, indicating that the user has the behavior category. For example, there is a directed edge between user A and the "payment record" behavior node, indicating that user A has a payment record.

[0078] In addition, there are undirected edges between different behavior nodes, indicating that there is a certain correlation between different behavior categories. For example, there is an undirected edge between the two behavior nodes "payment record" and "power load curve", indicating that there may be a certain correlation between the user's payment behavior and power load.

[0079] In the implementation of this article, the following three typical user behaviors are considered:

[0080] Payment records: including the user's historical payment time, amount, method, etc. Power load curve: reflects the changes in the user's power consumption in different time periods. Interactive service log: records the interaction between the user and the power company's customer service system, such as consultation, complaints, repairs, etc.

[0081] In the constructed user-behavior bipartite graph, it is necessary to reasonably estimate the weight of the associated edges between different behavior nodes. This paper considers the following factors to calculate the associated edge weight:

[0082] Behavior co-occurrence frequency: Counts the frequency of two behavior categories appearing simultaneously in a user set. The higher the frequency, the stronger the correlation between the two behaviors.

[0083] Total number of users: Normalize the co-occurrence frequency to reduce the impact of differences in the number of users.

[0084] Behavior temporal similarity: For certain time-related behaviors (such as payment records and electricity load curves), consider their temporal similarity. The closer the behaviors are in time, the stronger the correlation is.

[0085] Balance factor: A parameter is introduced to balance the influence of co-occurrence frequency and temporal similarity.

[0086] For example, assume that behavior A and behavior B appear together 500 times in 1,000 users, the total number of users is 10,000, the temporal similarity of the two behaviors is 0.8, and the balance factor is set to 0.6. The weight of the associated edge between A and B is calculated as follows:

[0087] The weight of the associated edge = 0.6*(500 / 10000)+0.4*0.8=0.35;

[0088] On the constructed user-behavior bipartite graph, a multi-layer graph attention network (GAT) is designed to learn the high-level semantic representation of user nodes and behavior nodes.

[0089] Each layer of GAT consists of a graph convolution layer and an attention layer. The graph convolution layer is responsible for propagating node feature information on the graph structure so that each node aggregates the features of its neighboring nodes. The attention layer adaptively calculates the importance weights of different neighboring nodes by introducing the attention mechanism.

[0090] Taking the l-th layer GAT as an example, for node i, the update process of its hidden state is as follows:

[0091] The graph convolution layer aggregates the hidden states of neighbor nodes according to the neighbor node set N(i) of node i to obtain the intermediate representation of node i:

[0092]

[0093] Among them, W is a learnable weight matrix, Aggregate represents an aggregation function (such as summation or averaging), and ReLU is an activation function.

[0094] The attention layer calculates the attention weight between node i and its neighbor node j:

[0095] α_ij=Softmax(LeakyReLU(a·Concat(W·h_i^l,W·h_j^l)));

[0096] Among them, a is a learnable attention weight vector, LeakyReLU is the activation function, and Softmax is used to normalize the attention weights.

[0097] Apply the attention weights to the hidden states of the neighboring nodes to get the final updated representation of node i:

[0098] h_i^(l+1)=ReLU(∑_{j∈N(i)}α_ij·(W·h_j^l));

[0099] Through iterative updates of multi-layer GAT, high-level semantic representations of user nodes and behavior nodes can be obtained.

[0100] After obtaining the high-level representation of user nodes and behavior nodes, a behavior feature fusion layer is designed to aggregate multiple behavior node information associated with users and generate a comprehensive user behavior feature representation vector.

[0101] Specifically, the attention pooling mechanism is used to adaptively assign weights to different behavior nodes according to the association strength between the behavior nodes and users. The behavior nodes with greater association strength should be given higher weights during fusion.

[0102] For example, assuming that the behavior nodes associated with user A include "payment record", "power load curve" and "interactive service log", their high-level representations are h_1, h_2, h_3, and their association strengths with user A are 0.8, 0.6, and 0.4, respectively. The behavior feature representation vector h_A of user A is calculated as follows:

[0103] h_A=0.8*h_1+0.6*h_2+0.4*h_3;

[0104] Finally, the user's behavior feature representation vector and the user's attribute feature representation vector are concatenated and fused to obtain the user's multimodal feature representation as the final user portrait vector.

[0105] User attribute features include static attribute information of users, such as age, gender, region, power customer type, etc. These attribute features can be directly extracted from the user feature dataset and converted into vector representation.

[0106] Assume that the attribute feature representation vector of user A is v_A, then the multimodal feature representation u_A of user A is calculated as follows:

[0107] u_A=Concat(v_A,h_A);

[0108] Among them, Concat represents a vector concatenation operation.

[0109] Through the above steps, the multimodal user feature representation learning model in this paper can fully mine and integrate the user's feature information in different modes, characterize the user's characteristics from the perspectives of static attributes and dynamic behaviors, and generate a comprehensive and accurate user portrait vector representation. This lays the foundation for the subsequent generation of personalized outbound call strategies and user response prediction.

[0110] In an optional embodiment,

[0111] By using knowledge graph embedding technology, the user portrait vector is mapped to the semantic space of the knowledge graph in the power field, realizing the deep integration of user characteristics and power business knowledge, including:

[0112] Construct a time-series dynamic knowledge graph, discretize the continuous time axis into a series of time slices, divide entities, relationships, and triples into corresponding time slices according to the timestamp of the data, construct a static knowledge subgraph on each time slice, and add time-series evolution edges between adjacent time slices to describe the change rules of entities and relationships over time, forming a complete time-series dynamic knowledge graph;

[0113] In each time slice, the TransE model considering time attenuation is used to optimize the time-aware embedding loss function. By minimizing the time-aware embedding loss function in each time slice, the time-related entity embedding representation is obtained. At the same time, the time-smoothing regularization term is introduced to consider the smoothness constraints between adjacent time slices to ensure the continuity of entity and relationship embedding in the time dimension. By jointly optimizing the embedding loss function and time-smoothing constraints of each time slice, the time-series knowledge representation is obtained, which not only retains the static semantic information of each time slice, but also obtains the dynamic evolution law across time slices.

[0114] Given a user's multimodal feature representation at a specific time slice, it is converted into a knowledge embedding space through nonlinear mapping, and the attention weights between the user features and each entity embedding are calculated to indicate the user's attention to different entities.

[0115] Introducing the temporal attention mechanism in the time dimension, calculating the temporal weight of each time slice, dynamically adjusting the knowledge weights of different time slices, comprehensively considering the entity attention weight and the temporal attention weight, enhancing the temporal knowledge representation, and obtaining the user's temporal knowledge enhanced feature representation;

[0116] The time series knowledge enhanced feature representation is spliced ​​with user features to obtain the final dynamic user representation, realizing the deep integration of user features and power business knowledge.

[0117] For example, after obtaining the multimodal user feature representation, in order to further improve the semantic expression ability of the user portrait, this paper uses the knowledge graph embedding technology to map the user features to the semantic space of the power field knowledge graph, and realizes the deep integration of user features and power business knowledge. The specific method steps are as follows:

[0118] Knowledge graphs in the power sector usually contain a lot of time-related dynamic information, such as electricity price policy adjustments, equipment status changes, etc. In order to characterize these dynamic characteristics, this paper constructs a time-series dynamic knowledge graph.

[0119] First, the continuous time axis is discretized into a series of time slices, each of which represents a fixed time interval, such as a month or a quarter. Then, entities, relations, and triples are divided into corresponding time slices according to the timestamp information of triples in the knowledge graph.

[0120] In each time slice, a static knowledge subgraph is constructed based on the divided entities and relationships. At the same time, temporal evolution edges are added between adjacent time slices to characterize the changing patterns of entities and relationships over time.

[0121] For example, suppose the knowledge graph contains the following triples:

[0122] (User A, payment method, bank deduction)-Timestamp: January 2022;

[0123] (User A, payment method, mobile payment)-Timestamp: April 2022;

[0124] (User B, payment method, bank deduction)-Timestamp: January 2022;

[0125] Dividing the time slices by quarter, we can get two time slices: the first quarter of 2022 and the second quarter of 2022. In the static knowledge subgraph of the first quarter, there are two triples (user A, payment method, bank deduction) and (user B, payment method, bank deduction). In the static knowledge subgraph of the second quarter, there is a triple (user A, payment method, mobile payment). At the same time, a time evolution edge is added between the two time slices, indicating that user A's payment method changes from bank deduction to mobile payment.

[0126] Through the above method, a complete temporal dynamic knowledge graph is constructed, which not only retains the static semantic information of each time slice, but also describes the dynamic evolution rules across time slices.

[0127] In order to obtain high-quality knowledge representation, this paper uses the TransE model considering time decay on each time slice to optimize the timing-aware embedding loss function.

[0128] The TransE model is based on the translation invariance assumption, that is, for a triple (h, r, t), the head entity h should be close to the tail entity t after translation by relation r. At each time slice, the embedding loss function is defined according to this assumption, and the embedding representation of entities and relations is learned by minimizing the loss function.

[0129] At the same time, in order to consider the time decay effect, the loss function is modified. Specifically, a time decay factor is introduced to give smaller weights to triples that are farther away from the current time slice, indicating that their contribution to the knowledge representation at the current moment is smaller.

[0130] In addition, in order to ensure the continuity of entity and relationship embedding in the time dimension, a temporal smoothing regularization term is introduced. For the same entity or relationship in adjacent time slices, its embedding representation should be as close as possible. The temporal smoothing regularization term is achieved by minimizing the difference in the embedding representation of adjacent time slices.

[0131] By jointly optimizing the embedding loss function and temporal smoothing constraints of each time slice, a high-quality temporal knowledge representation is obtained, which not only retains the static semantic information of each time slice, but also captures the dynamic evolution laws across time slices.

[0132] For example, assuming that at two adjacent time slices t1 and t2, the embedding representations of the entity "user A" are e_A^t1 and e_A^t2 respectively, then the temporal smoothing regularization term can be defined as:

[0133] R_smooth=||e_A^t1-e_A^t2||_2^2;

[0134] Among them, ||·||_2 represents the L2 norm. Minimizing this regularization term can make the embedding representation of user A in adjacent time slices as close as possible, ensuring the continuity in the time dimension.

[0135] Given a user's multimodal feature representation u^t at a specific time slice t, it is converted to the knowledge embedding space through nonlinear mapping to obtain the user's representation u_e^t in the knowledge space:

[0136] u_e^t=tanh(W·u^t+b);

[0137] Among them, W and b are the learnable mapping matrix and bias vector, and tanh is the hyperbolic tangent activation function.

[0138] Then, the attention weight a_i^t between the user representation u_e^t and each entity embedding e_i^t at time slice t is calculated:

[0139] a_i^t=softmax(u_e^t·e_i^t);

[0140] The attention weight a_i^t indicates the user's attention to entity i. The larger the weight, the stronger the association between the user and the entity.

[0141] For example, assuming that at time slice t, the knowledge space of user A is represented as u_A^t, and the embedding representations of entities "bank deduction" and "mobile payment" are e_bank^t and e_wechat^t respectively, then the attention weights of user A and these two entities are calculated as follows:

[0142] a_bank^t=exp(u_A^t·e_bank^t) / (exp(u_A^t·e_bank^t)+exp(u_A^t·e_wechat^t));

[0143] a_wechat^t=exp(u_A^t·e_wechat^t) / (exp(u_A^t·e_bank^t)+exp(u_A^t·e_wechat^t));

[0144] If user A mainly uses bank deduction to pay bills in time slice t, the value of a_bank^t will be larger, indicating that the user pays more attention to the "bank deduction" entity.

[0145] After obtaining the attention weights of user-entity interactions in the previous step, the temporal attention mechanism is further introduced to dynamically adjust the knowledge weights of different time slices.

[0146] Specifically, for each time slice t, calculate its temporal attention weight β^t:

[0147] β^t=softmax(v·tanh(W_t·u^t+b_t));

[0148] Where v is a learnable temporal attention vector, W_t and b_t are the mapping matrix and bias vector corresponding to time slice t. The temporal attention weight β^t represents the importance of time slice t to the user's current state.

[0149] Taking into account the entity attention weight a_i^t and the temporal attention weight β^t, we can obtain the knowledge-enhanced feature representation u_k^t of the user at time slice t:

[0150] u_k^t=∑_i(a_i^t·e_i^t)·β^t;

[0151] Finally, the knowledge-enhanced feature representations of each time slice are weighted and summed to obtain the user's final temporal knowledge-enhanced feature representation u_k:

[0152] u_k=∑_t u_k^t;

[0153] For example, suppose there are two time slices t1 and t2, and the multimodal feature representations of user A in these two time slices are u_A^t1 and u_A^t2 respectively. According to step 3, the attention weights of user A and the entities "bank deduction" and "mobile payment" in each time slice are calculated as a_bank^t1, a_wechat^t1, a_bank^t2, a_wechat^t2 respectively.

[0154] According to the temporal attention mechanism, the temporal weights β^t1 and β^t2 of time slices t1 and t2 are calculated. Assume that β^t1 = 0.7 and β^t2 = 0.3, which means that time slice t1 is more important to the user's current state.

[0155] Then the knowledge enhancement feature representations of user A at time slices t1 and t2 are:

[0156] u_A_k^t1=(a_bank^t1·e_bank^t1+a_wechat^t1·e_wechat^t1)·0.7;

[0157] u_A_k^t2=(a_bank^t2·e_bank^t2+a_wechat^t2·e_wechat^t2)·0.3;

[0158] Finally, the temporal knowledge enhancement feature of user A is expressed as:

[0159] u_A_k=u_A_k^t1+u_A_k^t2;

[0160] In order to obtain the final dynamic user representation, the temporal knowledge enhanced feature representation u_k is concatenated with the user's multimodal feature representation u to obtain the fused user representation u_final;

[0161] Through the above steps, this paper uses knowledge graph embedding technology to map user features to the semantic space of the knowledge graph in the power field, and realizes the deep integration of user features and power business knowledge through the attention mechanism. The fused dynamic user representation not only contains the user's static attributes and dynamic behavior characteristics, but also incorporates the background knowledge of the power field. It has stronger semantic expression ability and interpretability, and provides richer and more accurate user portrait information for the subsequent generation of personalized outbound call strategies.

[0162] In an optional embodiment,

[0163] The joint optimization of the embedding loss function and the temporal smoothing constraints of each time slice includes:

[0164]

[0165] Where L represents the loss value at the kth time step, (h, r, t)∈T t represents the true triple set T at the tth time step t A triple in , where h represents the head entity, r represents the relationship, and t represents the tail entity, (h′, r, t′)∈T′ t Represents the negative sampling triple set T′ at the tth time step t A triple in , where h′ represents the head entity of negative sampling, t′ represents the tail entity of negative sampling, y represents the margin parameter used to control the interval between positive and negative triplets, d() represents the distance function, and w k represents the time weight vector of the kth time step, [] + represents the operation of taking the positive part, (h+r)⊙w k represents the element-wise multiplication of the head entity representation under relation r with the temporal weight vector to obtain the time-aware head entity representation, t⊙w krepresents the element-wise multiplication of the embedding representation of the tail entity with the temporal weight vector to obtain the time-aware tail entity representation, (h′+r)⊙w k It means to multiply the representation of the negative sampling head entity under the relation r by the time weight vector at the element level to obtain the time-aware negative sampling head entity, t′⊙w k It means to perform element-wise multiplication of the embedding representation of the negative sampling tail entity with the time weight vector to obtain the time-aware negative sampling tail entity;

[0166]

[0167] Among them, R represents the regularization term, K represents the total number of time steps, |E k | represents the entity set E at the kth time step k The size of represents the embedding representation of the i-th entity at the k-th time step, Represents the embedding representation of the i-th relation at the k-th time step.

[0168] In an optional embodiment,

[0169] Combined with the outbound call business scenario, the attention mechanism and policy gradient algorithm are used to autonomously learn and generate personalized outbound call strategies for different user groups, and the personalized outbound call strategies are simulated, evaluated and dynamically optimized, including:

[0170] Define a set of outbound call strategy elements, including outbound call time, frequency, and content. Each strategy element has multiple optional values ​​that constitute its value space. For a given user, its personalized outbound call strategy is defined as a joint probability distribution over a set of strategy elements, which indicates the probability of adopting a specific combination of strategy elements given the user's characteristics.

[0171] Construct a personalized strategy generation model based on the attention mechanism. Take user features as input, dynamically associate user features with strategy elements through the attention mechanism, and use the attention mechanism to calculate the conditional probability distribution of each strategy element under given user features. Use the attention scoring function to measure the correlation between user features and strategy elements. According to different user features, dynamically adjust the importance weight of each strategy element to generate a personalized strategy distribution.

[0172] In the model training phase, the maximum likelihood estimation method is used to minimize the negative log-likelihood loss function. The attention parameters are learned through the stochastic gradient descent optimization algorithm, so that the personalized strategy generation model can generate personalized outbound call strategies that meet the user's preferences based on the user's characteristics.

[0173] The policy gradient algorithm of reinforcement learning is introduced. Through trial and error learning in a simulation environment, each outbound call attempt is defined as a sequential decision-making process interacting with the user. The state is the user characteristics at the current moment and the information of historical outbound call records. The action is the outbound call strategy sampled from the strategy distribution. The reward is the effect evaluation of the outbound call. The outbound call trajectory data is collected by interacting with the environment.

[0174] According to the policy gradient theorem, the parameters of the personalized strategy generation model are updated along the gradient direction to maximize the expected cumulative reward. The state-action value function is approximately estimated using the Monte Carlo method. The estimated value is substituted into the policy gradient formula to obtain the update rule for the parameters of the personalized strategy generation model. The personalized strategy generation model is optimized based on the outbound call trajectory data.

[0175] For example, after obtaining the dynamic user representation that integrates the knowledge of power business, this paper further combines the outbound call business scenario, uses the attention mechanism and policy gradient algorithm to autonomously learn and generate personalized outbound call strategies for different user groups, and simulates and evaluates and dynamically optimizes the generated strategies. The specific method steps are as follows:

[0176] First, define a set of outbound call strategy elements, including outbound call time, frequency, and content of the call. Each strategy element has multiple optional values, which constitute its value space. For example, outbound call time can be divided into time periods such as weekday mornings, weekday afternoons, and non-working days; outbound call frequency can be divided into frequencies such as once a day, once a week, and once a month; and the content of the call can be divided into types such as collection calls, marketing calls, and customer care calls.

[0177] Given a user, the personalized outbound calling strategy is defined as the joint probability distribution over this set of strategy elements, which represents the probability of adopting a specific combination of strategy elements given the user's characteristics. For example, for a user who travels frequently, the personalized outbound calling strategy may be to make outbound calls using customer care scripts on non-working days, with a frequency of once a month.

[0178] A personalized strategy generation model based on attention mechanism is constructed. The dynamic user representation u_final that integrates power business knowledge is used as input. The user characteristics and outbound call strategy elements are dynamically associated through the attention mechanism to generate personalized outbound call strategy distribution.

[0179] For each policy element a_i, the attention mechanism is used to calculate its conditional probability distribution p(a_i|u_final) given the user feature u_final:

[0180] p(a_i|u_final)=softmax(att_score(u_final,e_i));

[0181] Among them, e_i is the embedded representation of the strategy element a_i, and att_score is the attention scoring function used to measure the correlation between user features and strategy elements. Commonly used attention scoring functions include dot product, concatenation, etc.

[0182] Depending on the user characteristics, the attention mechanism can dynamically adjust the importance weights of various strategy elements. For example, for a user with a large amount of electricity bill arrears, the weight of the collection speech may be higher; while for a high-quality customer who has paid on time for a long time, the weight of the customer care speech may be higher.

[0183] Through the attention mechanism, the personalized strategy generation model can generate personalized outbound call strategy distribution that meets the preferences of different users based on their characteristics.

[0184] During the model training phase, the maximum likelihood estimation method is used to minimize the negative log-likelihood loss function. The parameters of the attention mechanism are learned through the stochastic gradient descent optimization algorithm, so that the personalized strategy generation model can generate personalized outbound call strategies that meet the user's preferences based on the user's characteristics.

[0185] Specifically, for a training sample (u_final, a), where u_final is a dynamic user representation that integrates power business knowledge, and a is the outbound call strategy actually accepted by the user, the negative log-likelihood loss function is defined as:

[0186] L(θ)=-log p(a|u_final;θ);

[0187] Among them, θ represents the parameters of the personalized strategy generation model, including the parameters of the attention mechanism.

[0188] Through the stochastic gradient descent algorithm, the model parameters θ are continuously updated iteratively to minimize the loss function L(θ), so that the model can learn the correlation between user characteristics and outbound call strategies and generate personalized outbound call strategy distribution.

[0189] For example, suppose there is a user sample (u_final, a), where u_final is a dynamic user representation that integrates electricity business knowledge, and a is the outbound call strategy actually accepted by the user, including the outbound call time being weekday afternoons, the outbound call frequency being once a week, and the content of the call being a collection call.

[0190] Based on the personalized strategy generation model, the conditional probability distribution of the user on each strategy element is calculated:

[0191] - Outbound call time: p(weekday morning|u_final) = 0.2, p(weekday afternoon|u_final) = 0.6, p(non-workday|u_final) = 0.2;

[0192] - Outbound call frequency: p(once a day|u_final) = 0.1, p(once a week|u_final) = 0.7, p(once a month|u_final) = 0.2;

[0193] - Content of the speech: p(collection speech|u_final)=0.8, p(marketing speech|u_final)=0.1, p(customer care speech|u_final)=0.1;

[0194] Then the negative log-likelihood loss of this sample is:

[0195] L(θ)=-log(0.6×0.7×0.8)=0.98;

[0196] By minimizing the loss function, the personalized strategy generation model can learn the user's preferences and make weekly outbound calls using collection language on weekday afternoons.

[0197] In order to further optimize the personalized strategy generation model and improve the effectiveness of the outbound call strategy, this paper introduces the policy gradient algorithm of reinforcement learning and continuously improves the strategy generation model through trial and error learning in a simulation environment.

[0198] Consider each outbound call attempt as a sequential decision-making process of interacting with the user, defined as follows:

[0199] - Status: current user characteristics, historical outbound call records, and other information.

[0200] -Action a: According to the outbound call strategy distribution generated by the personalized strategy generation model, sample and obtain specific outbound call strategies.

[0201] - Reward r: Evaluation of the effectiveness of the outbound call, such as whether payment collection was successful, whether user satisfaction was improved, etc.

[0202] By interacting with the simulation environment, a series of outbound call trajectory data (s, a, r) ​​are collected to optimize the personalized strategy generation model.

[0203] According to the policy gradient theorem, the gradient of the parameter θ of the personalized strategy generation model can be expressed as:

[0204]

[0205] Among them, π_θ is the policy distribution parameterized by the personalized policy generation model, Q^π(s_t,a_t) is the value function of the state-action pair (s_t,a_t) under the strategy π, which represents the expected cumulative reward starting from the state s_t, taking action a_t and continuing to interact according to the strategy π.

[0206] Since the value function Q^π cannot be obtained directly, the Monte Carlo method is needed for approximate estimation:

[0207] Q^π(s_t,a_t)≈∑_{t'=t}^Tγ^(t'-t)r_t'

[0208] Where γ is the discount factor and r_t' is the reward obtained at time t'.

[0209] Substituting the estimated value function into the policy gradient formula, we can obtain the update rule for the personalized policy generation model parameter θ:

[0210]

[0211] Here, α is the learning rate.

[0212] Based on the collected outbound call trajectory data, the above-mentioned update rules are used to continuously optimize the personalized strategy generation model so that it can generate more effective outbound call strategies.

[0213] For example, suppose an outbound call trace is collected:

[0214] -State s_0: User characteristics include a high amount of outstanding payments and no response to outbound calls in the past.

[0215] -Action a_0: The sampling outbound calling strategy is to use collection calls in the morning on weekdays.

[0216] - Reward r_0: If the user does not answer the call, the reward is -1.

[0217] -State s_1: User characteristics include high arrears, no response to past outbound calls, and no answer to the most recent outbound call.

[0218] -Action a_1: The sampling outbound calling strategy is to use collection calls in the afternoon on weekdays.

[0219] - Reward r_1: The user answers the call and promises to repay, and the reward is +5.

[0220] Based on the trajectory data, the estimated value function is:

[0221] -Q^π(s_0,a_0)≈r_0+γr_1=-1+0.9×5=3.5;

[0222] -Q^π(s_1,a_1)≈r_1=5;

[0223] Then the parameter update amount of the personalized strategy generation model is:

[0224]

[0225] By continuously collecting outbound call trajectory data and using the policy gradient algorithm to update model parameters, the personalized strategy generation model can be gradually optimized to learn more effective outbound call strategies and improve outbound call results.

[0226] After generating personalized outbound call strategies, this paper constructs a simulation environment to evaluate and dynamically optimize the strategies.

[0227] First, based on historical user data, a user behavior model is built to simulate user responses to different outbound call strategies, such as answering probability, commitment repayment probability, etc. Then, an outbound call strategy is generated using a personalized strategy generation model, which interacts with the user behavior model in a simulation environment, evaluates the effect of the outbound call strategy, and collects outbound call trajectory data.

[0228] Based on the collected outbound call trajectory data, the policy gradient algorithm is used to continuously optimize the personalized strategy generation model to generate more effective outbound call strategies. At the same time, through the evaluation feedback of the simulation environment, the outbound call strategy is dynamically adjusted, such as optimizing the outbound call time, frequency, and content of the words, to improve the outbound call success rate and user satisfaction.

[0229] Through simulation evaluation and dynamic optimization, personalized outbound calling strategies can be continuously improved to adapt to the preferences and feedback of different users, thereby improving the effectiveness and benefits of the overall outbound calling business.

[0230] For example, suppose a user behavior model is built to simulate the user's response to different outbound calling strategies:

[0231] -On weekday mornings, when making outbound calls, the probability of a user answering the call is 0.3, and the probability of a user promising to repay is 0.1.

[0232] -On weekday afternoons, the probability of a user answering the call is 0.6, and the probability of a user committing to repaying the loan is 0.4.

[0233] - When calling on non-working days, the probability of a user answering the call is 0.8 and the probability of a user committing to repaying the loan is 0.2.

[0234] Use the personalized strategy generation model to generate an outbound call strategy: make outbound calls using collection calls on weekday afternoons, interact with the user behavior model in a simulation environment, and evaluate the effectiveness of the strategy.

[0235] Assume that the simulation results are: under this strategy, the average user answering rate is 0.6, the average committed repayment rate is 0.4, and the outbound call success rate is 0.24.

[0236] According to the simulation evaluation results, the personalized strategy generation model can be further optimized, such as adjusting the outbound call time to a combination of weekday afternoons and non-working days to increase the user answering rate; optimizing the collection script content to increase the promised repayment rate.

[0237] Through continuous simulation evaluation and dynamic optimization, personalized outbound calling strategies can be continuously improved to enhance the overall effectiveness and benefits of outbound calling services.

[0238] In summary, this paper proposes a personalized outbound call strategy generation method based on reinforcement learning. Through the attention mechanism and policy gradient algorithm, it autonomously learns and generates personalized outbound call strategies for different user groups, and continuously improves the effect and benefits of outbound call strategies through simulation evaluation and dynamic optimization. This method makes full use of user characteristics and power business knowledge, can adapt to the preferences and feedback of different users, and provides a new idea and implementation solution for the intelligent outbound call system.

[0239] In an optional embodiment,

[0240] The personalized outbound call strategy is converted into a structured outbound call task script and sent to the intelligent outbound call platform. The intelligent outbound call platform uses domain ontology and semantic parsing technology to build a spoken dialogue generation model for power business, and automatically generates semantically coherent and logically self-consistent outbound call scripts, including:

[0241] Convert the personalized outbound call strategy into a structured task script. The task script adopts a hierarchical structure and contains information about the target user, outbound call time, script template, and business process of the outbound call task. The script template defines the overall structure and slots of the outbound call script, and the business process defines the business process and corresponding API interface that need to be executed during the outbound call process.

[0242] Construct an ontology knowledge base covering electricity consumption, payment, and energy conservation in the power sector. The ontology knowledge base contains the core concepts, attributes, and relationships in the field.

[0243] Perform semantic analysis on outbound call task scripts, identify key entities, attributes, and intentions in the scripts, map them to corresponding concepts in the ontology knowledge base, and associate corresponding attributes and measures;

[0244] A spoken dialogue generation model is constructed using a pre-trained language model. The domain ontology and semantic parsing results are used as input. A template-based generation method is adopted, and the fixed speech fragments in the task script are used as the template skeleton. The language model is used to dynamically fill in the slots in the template to generate complete outbound call speech that is semantically coherent and logically self-consistent.

[0245] For example, after generating a personalized outbound call strategy, in order to facilitate the intelligent outbound call system to execute the strategy, it is necessary to convert the strategy into a structured form and automatically generate the corresponding outbound call script. This paper proposes a method based on domain ontology and semantic parsing to convert personalized outbound call strategies into structured outbound call task scripts, and use the spoken dialogue generation model to automatically generate semantically coherent and logically self-consistent outbound call scripts. The specific steps of the method are as follows:

[0246] First, the personalized outbound call strategy is converted into a structured task script. The task script adopts a hierarchical structure and includes various elements and processes of the outbound call task.

[0247] Specifically, the task script contains the following information:

[0248] -Target users: target users of outbound call tasks, including user ID, user characteristics, etc.

[0249] -Outbound call time: the execution time of the outbound call task, such as weekday morning, weekday afternoon, etc.

[0250] - Script template: The overall structure and slots of the outbound call script define the framework of the script.

[0251] -Business process: The business process that needs to be executed during the outbound call, such as querying arrears information, generating payment orders, etc., as well as the corresponding API interface.

[0252] For example, assume that the personalized outbound call strategy is: for users with high outstanding balances, collection calls are made in the afternoon on weekdays, with a frequency of once a week.

[0253] In this task script, the ID and characteristics of the target user are defined, the outbound call time is weekday afternoons, and the frequency is once a week. The script template defines the overall structure of the outbound call script, including greetings, overdue payment notifications, payment persuasion, payment method guidance, and closing remarks. {overdue_amount}, {payment_methods}, etc. are slots that need to be filled dynamically. The business process defines the operations that need to be performed during the outbound call process, such as querying overdue payments and generating payment orders, as well as the corresponding API interfaces and parameters.

[0254] In order to generate semantically coherent outbound call scripts that conform to the logic of power business, it is necessary to build an ontology knowledge base covering knowledge in the power field.

[0255] The ontology knowledge base contains the core concepts, attributes and relationships in the power field, covering electricity consumption, payment, energy saving and other aspects. For example, the core concepts may include "user", "meter", "electricity fee", "payment method", etc.; attributes may include "user ID", "amount of outstanding fees", "payment status", etc.; relationships may include "user-owns-meter", "meter-generates-electricity fee", "user-pays-electricity fee", etc.

[0256] By constructing an ontology knowledge base, knowledge in the power field can be organized and represented in a structured manner, providing support for subsequent semantic parsing and speech generation.

[0257] For example, suppose a simple ontology knowledge base in the power field is built, which contains the following contents:

[0258] -Concepts: users, electricity meters, electricity charges, and payment methods;

[0259] -property:

[0260] -User: user ID, electricity address, contact number;

[0261] -Electricity meter: meter ID, user, current reading;

[0262] -Electricity bill: amount owed, payment status, payment deadline;

[0263] - Payment method: method name, operation process;

[0264] -relation:

[0265] -User-owned-meter;

[0266] -Electricity meter - generates -electricity charges;

[0267] -Users pay electricity bills;

[0268] -Electricity bills-Payment by-Payment method;

[0269] The ontology knowledge base represents the core concepts in the power field and the relationships between them in a structured way, such as users own electricity meters, electricity meters generate electricity bills, and users pay electricity bills through payment methods, etc., providing domain knowledge support for semantic parsing and speech generation.

[0270] After obtaining the structured outbound call task script, it is necessary to perform semantic parsing on the script, identify the key entities, attributes and intentions therein, and map them to the corresponding concepts in the ontology knowledge base, and associate the corresponding attributes and measures.

[0271] Specifically, the semantic parsing process includes the following steps:

[0272] Entity recognition: Identify key entities in the task script, such as user ID, amount of outstanding fees, etc. Attribute association: Associate the identified entities with the corresponding attributes in the ontology knowledge base, such as user ID corresponds to the "user ID" attribute of the "user" concept. Intent recognition: Identify business intents in task scripts, such as querying outstanding fees, generating payment orders, etc. Concept mapping: Map the identified intents to the corresponding concepts and relationships in the ontology knowledge base, such as "querying outstanding fees" is mapped to the "electricity fee" concept and the "user-payment-electricity fee" relationship.

[0273] Through semantic analysis, key information in the outbound call task script is extracted and associated with the domain knowledge in the ontology knowledge base to provide semantic support for subsequent speech generation.

[0274] For example, for the above task script example, the semantic parsing result can be expressed as:

[0275] -entity:

[0276] - User ID: "1001";

[0277] -Overdue amount: "{overdue_amount}";

[0278] -Attribute association:

[0279] - "1001"->user.userid;

[0280] -"{overdue_amount}"->Electricity bill.Amount overdue;

[0281] -intention:

[0282] -"query_overdue"->Query overdue payments;

[0283] -"generate_order"->Generate payment order;

[0284] -Concept Mapping:

[0285] -"query_overdue"->electricity bill, user-payment-electricity bill;

[0286] -"generate_order"->Electricity bill, electricity bill-by-payment method-payment;

[0287] Through semantic analysis, the key entities "1001" and "{overdue_amount}" in the task script are identified and associated with the "user.user ID" and "electricity fee.overdue amount" attributes in the ontology knowledge base; the two intentions "query overdue fees" and "generate payment order" are identified and mapped to the "electricity fee" concept and the relationships "user-payment-electricity fee" and "electricity fee-through-payment method-payment". These semantic information provides important support for the subsequent speech generation.

[0288] After obtaining the semantic analysis results of the outbound call task script, a pre-trained language model is used to build a spoken dialogue generation model to automatically generate semantically coherent and logically self-consistent outbound call scripts.

[0289] The dialogue generation model takes domain ontology and semantic parsing results as input and adopts a template-based generation method. Specifically, the fixed speech fragments in the task script are used as the template skeleton, and the language model is used to dynamically fill in the slots in the template to generate a complete outbound call speech.

[0290] The generation process is as follows:

[0291] Selection of speech template: Based on the results of semantic analysis, select the appropriate speech template, such as the collection speech template.

[0292] Slot filling: Use the language model to dynamically fill in the slots in the template based on ontology knowledge and context information, such as replacing "{overdue_amount}" with the actual overdue amount.

[0293] Speech optimization: Optimize the generated speech, such as adjusting word order, replacing synonyms, etc., to improve the fluency and naturalness of the speech.

[0294] Logical consistency check: Check whether the generated script complies with the logic and business process of the task script to ensure the logical consistency of the script.

[0295] Through the above steps, using ontology knowledge and semantic analysis results, semantically coherent and logically self-consistent outbound call scripts are dynamically generated, realizing the automatic conversion of personalized outbound call strategies into specific scripts.

[0296] For example, for the above task script and semantic parsing results, the generated outbound call script can be:

[0297] "Hello, I am a customer service representative from the power company. According to our records, your current electricity bill arrears amount is 500 yuan. In order to avoid affecting your normal electricity use, please pay the electricity bill in time. You can pay through online banking, the power company's business hall, etc. Thank you for your cooperation, goodbye."

[0298] In this script, the collection script template was selected, and the "{overdue_amount}" slot was dynamically filled with the actual overdue amount "500 yuan". According to the payment method information in the ontology knowledge base, prompts for payment methods such as "online banking, power company business hall" were generated. At the same time, the script was optimized and the word order was adjusted to improve the fluency and naturalness of the script. Finally, the script was checked to see if it complied with the business processes such as querying overdue payments and generating payment orders in the task script to ensure the logical consistency of the script.

[0299] In summary, this paper proposes a method based on domain ontology and semantic parsing to transform personalized outbound call strategies into structured outbound call task scripts, and uses a spoken dialogue generation model to automatically generate semantically coherent and logically self-consistent outbound call scripts. This method makes full use of domain knowledge and semantic information, realizes the automatic transformation of outbound call strategies into scripts, improves the personalization and naturalness of outbound call scripts, and provides support for the application of intelligent outbound call systems.

[0300] Figure 2FIG. 1 is a schematic diagram of the structure of an intelligent outbound calling system for a power company based on artificial intelligence according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0301] The first unit is used to obtain user data of the power company, wherein the user data includes basic information of the user, electricity usage information, payment information, and interactive service records; and use data mining technology to perform data cleaning, data integration, data transformation, and data reduction on the user data to generate a standardized user feature data set;

[0302] The second unit is used to construct a multimodal user feature representation learning model for the user feature data set, construct a user-behavior bipartite graph through a graph neural network algorithm, determine the user's behavior feature representation vector and generate a user portrait vector that integrates multi-dimensional information; use knowledge graph embedding technology to map the user portrait vector to the semantic space of the knowledge graph in the power field, and realize the deep integration of user features and power business knowledge;

[0303] The third unit is used to combine the outbound call business scenario, use the attention mechanism and policy gradient algorithm, autonomously learn and generate personalized outbound call strategies for different user groups, and simulate, evaluate and dynamically optimize the personalized outbound call strategies; convert the personalized outbound call strategies into structured outbound call task scripts, and send them to the intelligent outbound call platform. The intelligent outbound call platform uses domain ontology and semantic parsing technology to build a spoken dialogue generation model for power business, and automatically generates semantically coherent and logically self-consistent outbound call scripts.

[0304] According to a third aspect of the embodiments of the present invention,

[0305] An electronic device is provided, comprising:

[0306] processor;

[0307] a memory for storing processor-executable instructions;

[0308] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0309] A fourth aspect of the embodiments of the present invention is:

[0310] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0311] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0312] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent outbound calling method for power companies based on artificial intelligence, characterized in that: include: Obtain user data from the power company, including basic information, electricity usage information, payment information, and interactive service records of the user; Using data mining technology to perform data cleaning, data integration, data transformation and data reduction on the user data to generate a standardized user feature data set; For the user feature data set, a multimodal user feature representation learning model is constructed. Through the graph neural network algorithm, a user-behavior bipartite graph is constructed to determine the user's behavior feature representation vector and generate a user portrait vector that integrates multi-dimensional information; using the knowledge graph embedding technology, the user portrait vector is mapped to the semantic space of the knowledge graph in the power field to achieve a deep integration of user features and power business knowledge; In combination with the outbound call business scenario, the attention mechanism and policy gradient algorithm are used to autonomously learn and generate personalized outbound call strategies for different user groups, and the personalized outbound call strategies are simulated, evaluated and dynamically optimized; the personalized outbound call strategies are converted into structured outbound call task scripts, and sent to the intelligent outbound call platform. The intelligent outbound call platform uses domain ontology and semantic parsing technology to build a spoken dialogue generation model for power business, and automatically generates semantically coherent and logically self-consistent outbound call scripts.

2. The method according to claim 1, characterized in that: For the user feature data set, a multimodal user feature representation learning model is constructed, and a user-behavior bipartite graph is constructed through a graph neural network algorithm to determine the user's behavior feature representation vector and generate a user portrait vector that integrates multi-dimensional information, including: Constructing a user-behavior bipartite graph, wherein the user-behavior bipartite graph includes user nodes and behavior nodes, and there are directed edges between user nodes and behavior nodes; there are undirected edges between behavior nodes, indicating that there is a correlation between different behavior categories; the behavior categories include user payment records, power load curves, and interactive service logs; Calculate the weight of the associated edge between the behavior nodes, comprehensively consider the co-occurrence frequency and time sequence pattern of the behavior, and determine the weight of the associated edge based on the co-occurrence frequency of the behavior in the user set, the total number of users, the similarity of the behavior in time, and the balance factor; A multi-layer graph attention neural network is designed on the user-behavior bipartite graph. The node hidden state is updated by aggregating neighbor node information, and the importance weights of different neighbor nodes are adaptively calculated through the attention mechanism to learn the high-level semantic representation of the user node and the behavior node. Each layer of the graph attention neural network includes a graph convolution layer and an attention layer. The graph convolution layer propagates node features through graph convolution, and the attention layer calculates the attention weights between nodes by introducing an attention weight vector. After obtaining high-level semantic representations of user nodes and behavior nodes through the graph attention neural network learning, a behavior feature fusion layer is designed, and an attention pooling mechanism is adopted to adaptively assign weights to different behavior nodes according to the strength of association between the behavior nodes and the user, and the hidden states of the behavior nodes are weighted and fused to obtain the user's behavior feature representation vector; The user behavior feature representation vector and the user attribute feature representation vector are concatenated and fused to construct a multimodal feature representation of the user as a user portrait vector, wherein the user attribute feature representation vector includes static attribute information of the user, including but not limited to age, gender, region, and electricity customer type.

3. The method according to claim 1, characterized in that By using knowledge graph embedding technology, the user portrait vector is mapped to the semantic space of the knowledge graph in the power field, realizing the deep integration of user characteristics and power business knowledge, including: Construct a time-series dynamic knowledge graph, discretize the continuous time axis into a series of time slices, divide entities, relationships, and triples into corresponding time slices according to the timestamp of the data, construct a static knowledge subgraph on each time slice, and add time-series evolution edges between adjacent time slices to describe the change rules of entities and relationships over time, forming a complete time-series dynamic knowledge graph; In each time slice, the TransE model considering time attenuation is used to optimize the time-aware embedding loss function. By minimizing the time-aware embedding loss function in each time slice, the time-related entity embedding representation is obtained. At the same time, the time-smoothing regularization term is introduced to consider the smoothness constraints between adjacent time slices to ensure the continuity of entity and relationship embedding in the time dimension. By jointly optimizing the embedding loss function and time-smoothing constraints of each time slice, the time-series knowledge representation is obtained, which not only retains the static semantic information of each time slice, but also obtains the dynamic evolution law across time slices. Given a user's multimodal feature representation at a specific time slice, it is converted into a knowledge embedding space through nonlinear mapping, and the attention weights between the user features and each entity embedding are calculated to indicate the user's attention to different entities. Introducing the temporal attention mechanism in the time dimension, calculating the temporal weight of each time slice, dynamically adjusting the knowledge weights of different time slices, comprehensively considering the entity attention weight and the temporal attention weight, enhancing the temporal knowledge representation, and obtaining the user's temporal knowledge enhanced feature representation; The time series knowledge enhanced feature representation is spliced ​​with user features to obtain the final dynamic user representation, realizing the deep integration of user features and power business knowledge.

4. The method according to claim 3, characterized in that The joint optimization of the embedding loss function and the temporal smoothing constraints of each time slice includes: Where L represents the loss value of the kth time step, (h, r, t)∈T t represents the true triple set T at the tth time step t A triple in , where h represents the head entity, r represents the relationship, and t represents the tail entity, (h′, r, t′)∈T′ t Represents the negative sampling triple set T′ at the tth time step t A triple in , where h′ represents the head entity of negative sampling, t′ represents the tail entity of negative sampling, y represents the margin parameter used to control the interval between positive and negative triplets, d() represents the distance function, and w k represents the time weight vector of the kth time step, [] + represents the operation of taking the positive part, (h+r)⊙w k represents the element-wise multiplication of the head entity representation under relation r with the temporal weight vector to obtain the time-aware head entity representation, t⊙w k It means to multiply the embedding representation of the tail entity by the time weight vector element-wise to obtain the time-aware tail entity representation, (h′+r)⊙w k It means to multiply the representation of the negative sampling head entity under the relation r by the time weight vector at the element level to obtain the time-aware negative sampling head entity, t′⊙w k It means to perform element-wise multiplication of the embedding representation of the negative sampling tail entity with the time weight vector to obtain the time-aware negative sampling tail entity; Among them, R represents the regularization term, K represents the total number of time steps, |E k | represents the entity set E at the kth time step k The size of represents the embedding representation of the i-th entity at the k-th time step, Represents the embedding representation of the i-th relation at the k-th time step.

5. The method according to claim 1, characterized in that Combined with the outbound call business scenario, the attention mechanism and policy gradient algorithm are used to autonomously learn and generate personalized outbound call strategies for different user groups, and the personalized outbound call strategies are simulated, evaluated and dynamically optimized, including: Define a set of outbound call strategy elements, including outbound call time, frequency, and content. Each strategy element has multiple optional values ​​that constitute its value space. For a given user, its personalized outbound call strategy is defined as a joint probability distribution over a set of strategy elements, which indicates the probability of adopting a specific combination of strategy elements given the user's characteristics. Construct a personalized strategy generation model based on the attention mechanism. Take user features as input, dynamically associate user features with strategy elements through the attention mechanism, and use the attention mechanism to calculate the conditional probability distribution of each strategy element under given user features. Use the attention scoring function to measure the correlation between user features and strategy elements. According to different user features, dynamically adjust the importance weight of each strategy element to generate a personalized strategy distribution. In the model training phase, the maximum likelihood estimation method is used to minimize the negative log-likelihood loss function. The attention parameters are learned through the stochastic gradient descent optimization algorithm, so that the personalized strategy generation model can generate personalized outbound call strategies that meet the user's preferences based on the user's characteristics. The policy gradient algorithm of reinforcement learning is introduced. Through trial and error learning in a simulation environment, each outbound call attempt is defined as a sequential decision-making process interacting with the user. The state is the user characteristics at the current moment and the information of historical outbound call records. The action is the outbound call strategy sampled from the strategy distribution. The reward is the effect evaluation of the outbound call. The outbound call trajectory data is collected by interacting with the environment. According to the policy gradient theorem, the parameters of the personalized strategy generation model are updated along the gradient direction to maximize the expected cumulative reward. The state-action value function is approximately estimated using the Monte Carlo method. The estimated value is substituted into the policy gradient formula to obtain the update rule for the parameters of the personalized strategy generation model. The personalized strategy generation model is optimized based on the outbound call trajectory data.

6. The method according to claim 1, characterized in that The personalized outbound call strategy is converted into a structured outbound call task script and sent to the intelligent outbound call platform. The intelligent outbound call platform uses domain ontology and semantic parsing technology to build a spoken dialogue generation model for power business, and automatically generates semantically coherent and logically self-consistent outbound call scripts, including: Convert the personalized outbound call strategy into a structured task script. The task script adopts a hierarchical structure and contains information about the target user, outbound call time, script template, and business process of the outbound call task. The script template defines the overall structure and slots of the outbound call script, and the business process defines the business process and corresponding API interface that need to be executed during the outbound call process. Construct an ontology knowledge base covering electricity consumption, payment, and energy conservation in the power sector. The ontology knowledge base contains the core concepts, attributes, and relationships in the field. Perform semantic analysis on outbound call task scripts, identify key entities, attributes, and intentions in the scripts, map them to corresponding concepts in the ontology knowledge base, and associate corresponding attributes and measures; A spoken dialogue generation model is constructed using a pre-trained language model. The domain ontology and semantic parsing results are used as input. A template-based generation method is adopted, and the fixed speech fragments in the task script are used as the template skeleton. The language model is used to dynamically fill in the slots in the template to generate complete outbound call speech that is semantically coherent and logically self-consistent.

7. An intelligent outbound call system for power companies based on artificial intelligence, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain user data of the power company, wherein the user data includes basic information of the user, electricity usage information, payment information, and interactive service records; and use data mining technology to perform data cleaning, data integration, data transformation, and data reduction on the user data to generate a standardized user feature data set; The second unit is used to construct a multimodal user feature representation learning model for the user feature data set, construct a user-behavior bipartite graph through a graph neural network algorithm, determine the user's behavior feature representation vector and generate a user portrait vector that integrates multi-dimensional information; use knowledge graph embedding technology to map the user portrait vector to the semantic space of the knowledge graph in the power field, and realize the deep integration of user features and power business knowledge; The third unit is used to combine the outbound call business scenario, use the attention mechanism and policy gradient algorithm, autonomously learn and generate personalized outbound call strategies for different user groups, and simulate, evaluate and dynamically optimize the personalized outbound call strategies; convert the personalized outbound call strategies into structured outbound call task scripts, and send them to the intelligent outbound call platform. The intelligent outbound call platform uses domain ontology and semantic parsing technology to build a spoken dialogue generation model for power business, and automatically generates semantically coherent and logically self-consistent outbound call scripts.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images