A method and device for allocating outbound call service cases

Abstract

The embodiment of the present application provides an outbound business case allocation method and device, which deeply mines the potential relationship between the combination of outbound business cases and seats and the results of completing the case based on a neural network model, and uses a deep neural network model to predict the probability of all case allocation combinations being able to complete the case goals. Based on the prediction results, a network flow graph is constructed and the optimal allocation combination between cases and seats is found, thereby achieving the allocation of suitable cases to suitable seats, which not only ensures the fairness of case allocation, but also improves the completion rate of case goals and is conducive to improving the work enthusiasm of the seats. It can be seen that the technical solution of the embodiment of the present application solves the problem that cases and seats in current outbound business scenarios such as collection or marketing cannot be optimally allocated.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method and device for allocating outbound call service cases. Background Art

[0002] In outbound business scenarios such as telemarketing or telephone collection, a batch of tasks usually includes many cases, which are generally assigned to several seats (such as salespersons in telemarketing business, debt collectors in telephone collection business, etc.) for processing. After receiving the case, the agent is responsible for calling the customer involved in the case and following up to achieve certain business goals (such as order conversion, collection, etc.). In outbound business scenarios, since different seats and different cases have their own characteristics, how to allocate cases and seats often has a great impact on the overall business results (such as the overall order conversion rate, remittance rate, etc.).

[0003] In the prior art, cases and seats are usually allocated in an indiscriminate manner. For example, a batch of cases is evenly allocated to multiple seats, or randomly allocated to multiple seats. This allocation method cannot achieve the optimal allocation between cases and seats based on the characteristics of the seats and the cases, and cannot maximize the overall profit. In addition, it is impossible to achieve a fair allocation of each seat in terms of the difficulty of collection, which is likely to reduce the enthusiasm of the seats and lead to poor business results. Summary of the invention

[0004] The embodiments of the present application provide a method and device for allocating outbound call service cases to solve the problem that cases and seats in current outbound call service scenarios such as collection or marketing cannot be optimally allocated.

[0005] In a first aspect, an embodiment of the present application provides a method for allocating outbound call service cases, including:

[0006] For each case, a probability estimate of the completion of the case target corresponding to each seat is generated respectively. The probability estimate of the completion of the case target corresponding to each seat includes: generating the current seat portrait information, the case called person portrait information, and the first convolution feature vector corresponding to the case information, and generating the second convolution feature vector corresponding to the called person's historical call content; cross-multiplying the first convolution feature vector and the second convolution feature vector to obtain a combined feature vector; pooling the combined feature vector using a variety of different pooling methods to obtain multiple intermediate representation vectors of the combined feature vector; concatenating the multiple intermediate representation vectors to obtain a final representation vector representing the case and the current seat; outputting the score of the final representation vector through an activation function, which is the probability estimate of the completion of the case target corresponding to the case and the current seat; and allocating seats to multiple cases according to the probability estimates corresponding to multiple cases and multiple seats.

[0007] In one implementation, a first convolution feature vector corresponding to the agent portrait information, the called party portrait information, and the case information of the case is generated, including: encoding the structured text of the agent portrait information, the called party portrait information, and the case information to obtain a first text feature vector; and using a convolutional neural network to extract convolution features from the first text feature vector to obtain a first convolution feature vector.

[0008] In one implementation, generating a second convolution feature vector corresponding to the called party's historical call content includes: converting the called party's historical call content into dialogue text information using speech synthesis technology; encoding the dialogue text information to obtain a corresponding second text feature vector; and processing the second text feature vector using a convolutional neural network to obtain a second convolution feature vector.

[0009] In one implementation, a convolutional neural network is used to extract convolutional features from the second text feature vector to obtain a second convolutional feature vector, including: using multiple convolution kernels of different sizes to extract convolutional features from the second text feature vector respectively, and splicing the extracted convolutional features to obtain the second convolutional feature vector.

[0010] In one implementation, the pooling methods include at least two of the following: maximum pooling, average pooling, global average pooling, mixed pooling, random pooling, and power average pooling.

[0011] In one implementation, the activation function is a sigmoid activation function.

[0012] In one implementation, seats are allocated to multiple cases based on probability estimates corresponding to multiple cases and multiple seats, including: taking each case in the case set and each seat in the seat set as a node, constructing a network flow graph N from the case set to the seat set, solving the minimum cost maximum flow of the network flow graph N to obtain the optimal solution for the allocation relationship between multiple seats and multiple cases; the network flow graph N = (V, E, c, w, s, t), where V represents a set of nodes, E represents a set of edges, the edge (vi, vj)∈E between any case i and any seat j, c is the capacity of the edge, w is the cost on the edge, the cost on the edge between any case i and any seat j w(vi, vj)=c-pij, pij is the probability estimate corresponding to any case i and any seat j, s is the origin of the network flow graph, and t is the sink of the network flow graph.

[0013] In one implementation, when the outbound call service case is a collection case, the agent portrait information is the collection agent portrait information, the called party portrait information is the collection object portrait information, and the case information includes at least the case amount and the number of overdue days.

[0014] In one implementation, when outbound call business cases are assigned in real time, the information used to generate the first convolution feature vector also includes the real-time status of the agent.

[0015] In a second aspect, an embodiment of the present application provides an outbound call service case allocation device, including: an input module, an encoding module, a combination and interaction module, and a feature aggregation and prediction module, which are used to generate, for each case, a probability estimate of the completion of the case target corresponding to each agent, wherein: the input module is used for current agent portrait information, case called portrait information, case information, and called party historical call content; the encoding module is used to generate a first convolution feature vector corresponding to the current agent portrait information, case called portrait information, and case information, and generate a second convolution feature vector corresponding to the called party historical call content; the combination and interaction module is used to combine the first convolution feature vector and the second convolution feature vector. The feature vectors are cross-multiplied to obtain a combined feature vector; the combination and interaction module is further used to use a variety of different pooling methods to perform pooling processing on the combined feature vector to obtain multiple intermediate representation vectors of the combined feature vector; the feature aggregation and prediction module is used to splice the multiple intermediate representation vectors to obtain a final representation vector representing the case and the current seat; the feature aggregation and prediction module is also used to output the score of the final representation vector through an activation function, and the score is the probability estimate of the completion of the case target corresponding to the case and the current seat; the case allocation device also includes: a solution module, which is used to allocate seats for multiple cases according to the probability estimates corresponding to multiple cases and multiple seats.

[0016] The technical solution provided in the embodiment of the present application deeply mines the potential relationship between the combination of outbound business cases and seats and the results of completing the case based on a neural network model, uses a deep neural network model to predict the probability of all case allocation combinations being able to complete the case goals, constructs a network flow graph based on the prediction results, and finds the optimal allocation combination between cases and seats, thereby achieving the allocation of suitable cases to suitable seats, which not only ensures the fairness of case allocation, but also improves the completion rate of case goals, and is beneficial to improving the work enthusiasm of the agents. It can be seen that the technical solution of the embodiment of the present application solves the problem that cases and seats in current outbound business scenarios such as collection or marketing cannot be optimally allocated. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flowchart of the overall idea of ​​the call service case allocation method provided in the embodiment of the present application;

[0018] Figure 2 is a flow chart of a method for allocating call service cases provided in an embodiment of the present application;

[0019] Figure 3 is a neural network model constructed in an embodiment of the present application;

[0020] Figure 4 is a flowchart of generating a first convolution feature vector provided in an embodiment of the present application;

[0021] Figure 5 is a flow chart of generating a second convolution feature vector provided in an embodiment of the present application;

[0022] Figure 6 is a schematic diagram of a network flow graph for solving a minimum cost maximum flow according to an embodiment of the present application;

[0023] Figure 7 It is a structural schematic diagram of an outbound call service case allocation device provided in an embodiment of the present application;

[0024] Figure 8 It is a structural diagram of an outbound call service case allocation device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to solve the problem that cases and seats in current outbound call business scenarios such as collection or marketing cannot be optimally allocated, an embodiment of the present application provides an outbound call business case allocation method.

[0026] Figure 1 This is a flowchart of the overall idea of ​​the call service case allocation method provided by the embodiment of the present application. Figure 1 As shown, taking the collection scenario as an example, the outbound call business case allocation method provided in the embodiment of the present application is generally implemented through two stages: "estimation of the possibility of collection" and "optimal matching solution".

[0027] In the "Collection Possibility Prediction" stage, the embodiment of the present application constructs a neural network model for collection rate estimation, and inputs the case's seat portrait information, called person portrait information, case information, etc. into the neural network model, which can calculate the probability estimate when any case is assigned to any collector, and then provide the probability estimate to the "Optimal Matching Solution" stage. In the "Optimal Matching Solution" stage, the embodiment of the present application uses the minimum cost maximum flow solution method to optimize the matching of cases and collectors based on the probability estimate obtained in the "Collection Possibility Prediction" stage to obtain the optimal allocation result of the case.

[0028] Figure 2 Flowchart of the method for allocating call service cases provided by the embodiment of the present application. Figure 2 As shown, the method may include the following steps S101-S106, wherein steps S101-S105 may correspond to Figure 1 In the "Estimation of the likelihood of payment collection" stage, step S106 can correspond to Figure 1 The "optimal matching solution" stage in .

[0029] Figure 3 is a neural network model constructed in the embodiment of the present application, and the neural network model can be used to implement steps S101 to S105 in the embodiment of the present application. In the collection scenario, the neural network model is used to generate a probability estimate of the completion of the case target corresponding to each seat for each collection case. Figure 3 As shown in the figure, the neural network model includes input layer, encoding layer, combination and depth interaction layer and feature aggregation and prediction layer. Figure 2 and Figure 3 An exemplary description is given of steps S101 to S105 of the embodiment of the present application, i.e., a method for generating a probability estimate of the completion of a case goal corresponding to each agent.

[0030] Step S101, generate the current agent portrait information, the case called portrait information, the first convolution feature vector corresponding to the case information, and generate the second convolution feature vector corresponding to the called party's historical call content.

[0031] In the embodiment of the present application, the agent portrait information refers to the information related to the agent itself in the outbound call business scenario, and the embodiment of the present application does not specifically limit this. For example, in the collection scenario, the agent is a collector, then the agent portrait information is the collector portrait information, and the collector portrait information may include: gender, age, historical collection recovery rate, repayment status of assigned cases, violation status, collector's working time, collection risk level, etc.

[0032] In the embodiment of the present application, the called party portrait information can be referred to as customer portrait information, which refers to information related to the called party itself in the outbound call business scenario, and the embodiment of the present application does not specifically limit this. In the collection scenario, the called party can be the collection object, and the called party portrait information is the collection object information; in the marketing scenario, the called party can be the marketing object, and the called party portrait information is the marketing object information. Exemplarily, in the collection scenario, the collection object information may include: gender, region, education level, income level, repayment willingness, repayment habits, attitude towards answering the phone, tendency to default, etc.

[0033] In the embodiment of the present application, case information is information related to the case itself in the outbound call business scenario, and the embodiment of the present application does not specifically limit this. In the collection scenario, case information may include, for example, the case amount, the number of days overdue, etc.; in the marketing scenario, case information may include, for example, the amount of goods, the number of days of contact, etc.

[0034] In an embodiment of the present application, the called party's historical call content may be the call content generated during the conversation between the outbound call robot and the called party before the case is assigned to a certain agent. The call content generally appears in the form of a dialogue.

[0035] In an embodiment of the present application, case allocation for outbound call services may be performed in real time or in non-real time. When real time allocation is performed, the information used to generate the first convolution feature vector may also include the real-time status of the agent.

[0036] In the embodiment of the present application, step S101 can be implemented in the input layer and the encoding layer of the neural network model. The following specifically describes the manner in which step S101 generates the first convolution vector and the second convolution vector.

[0037] Figure 4 is a flow chart of generating a first convolution feature vector provided by an embodiment of the present application. Figure 4 As shown, in one implementation, the step of generating a first convolution feature vector includes the following steps S201-S202:

[0038] Step S201, encode the structured text of the agent portrait information, the called party portrait information, and the case information to obtain a first text feature vector.

[0039] In the embodiment of the present application, step S201 can be implemented in the input layer of the neural network model, which can be, for example, a word embedding model. The agent portrait information, the called party portrait information, the case information, etc. are input into the input layer of the neural network model in the form of structured text, and the input layer encodes the agent portrait information, the called party portrait information, the case information, etc. into corresponding vector representations, that is, obtaining the first text feature vector.

[0040] Step S202: Use a convolutional neural network to extract convolution features from the first text feature vector to obtain a first convolution feature vector.

[0041] In an embodiment of the present application, step S202 can be implemented in the encoding layer of the neural network model, which encoding layer may include one or more convolutional neural networks. The convolutional neural network can perform a convolution operation on the first text feature vector input to the encoding layer to extract the convolution features of the first text feature vector and obtain a first convolution feature vector.

[0042] Figure 5 is a flow chart of generating a second convolution feature vector provided by an embodiment of the present application. Figure 5 As shown, in one implementation, the step of generating the second convolution feature vector includes the following steps S301-S303:

[0043] Step S301: convert the called party's historical call content into dialogue text information using speech synthesis technology.

[0044] Since the called party's historical call content is the content of the voice call between the called party and the outbound service robot, or the content of the called party's previous voice call with the agent, it is generally presented in the form of a recording. Therefore, the embodiment of the present application can use speech synthesis (Text To Speech, TTS) technology to convert the recording into text information, and the text information can be presented in the form of a dialogue between the called party and the outbound service robot, wherein the parts of the called party and the outbound service robot spoken by each other can be distinguished by tone recognition.

[0045] Step S302: Encode the dialogue text information to obtain a corresponding second text feature vector.

[0046] In the embodiment of the present application, step S302 can be implemented in the input layer of the neural network model. After the historical call content is converted into a form of dialogue text, it can be input into the input layer of the neural network model, and the input layer encodes the dialogue text into a corresponding vector representation, that is, obtaining a second text feature vector.

[0047] Step S303: Use a convolutional neural network to process the second text feature vector to obtain a second convolution feature vector.

[0048] In the embodiment of the present application, step S303 can be implemented in the coding layer of the neural network model. The convolutional neural network included in the coding layer can perform a convolution operation on the second text feature vector input to the coding layer to extract the convolution feature of the second text feature vector and obtain a second convolution feature vector.

[0049] In one implementation, the convolutional neural network can use multiple convolution kernels of different sizes to perform convolution operations on the second text feature vector respectively, thereby obtaining multiple convolution features, and then integrate the multiple convolution features through dimensional superposition, splicing or cross multiplication to obtain the second convolution feature vector.

[0050] Step S102, cross-multiply the first convolution feature vector and the second convolution feature vector to obtain a combined feature vector.

[0051] In the embodiment of the present application, step S102 can be implemented in the combination and deep interaction layer of the neural network model.

[0052] Exemplarily, assuming that the dimension of the first convolution feature vector is m and the dimension of the second convolution feature vector is n, after cross-producting the first convolution feature vector and the second convolution feature vector, the dimension of the combined feature vector that can be obtained is m×n.

[0053] In the embodiment of the present application, the first convolution feature vector and the second convolution feature vector are cross-multiplied to achieve the combination and deep interaction of the first convolution feature vector and the second convolution feature vector, so that the combined feature vector obtained after the interaction can reflect the intrinsic connection between the case and the seat.

[0054] Step S103: performing pooling processing on the combined feature vector using a plurality of different pooling methods to obtain a plurality of intermediate representation vectors of the combined feature vector.

[0055] In the embodiment of the present application, step S103 can be implemented in the combination and deep interaction layer of the neural network model.

[0056] It is understandable that since the combined feature vector is derived from the cross product of the first convolution feature vector and the second convolution feature vector, its dimension is relatively large. Therefore, if the combined feature vector is directly calculated, the model burden will be increased and there will be a risk of overfitting. Therefore, in step S103, the embodiment of the present application performs pooling on the combined feature vector, thereby increasing the receptive field of the application network, suppressing noise, reducing information redundancy, reducing the amount of model calculation, reducing the difficulty of network optimization, and preventing network overfitting. The pooling method used in the embodiment of the present application may include but is not limited to the following pooling methods: maximum pooling, average pooling, global average pooling, mixed pooling, random pooling, and power average pooling.

[0057] Take the combined feature vector m×n obtained in step S103 as an example:

[0058] Maximum pooling can be to divide the combined feature vector m×n into several sub-regions and output the maximum value for each sub-region.

[0059] Average pooling can be to divide the combined feature vector m×n into several sub-regions and output the average value for each sub-region.

[0060] Global average pooling can be performed by averaging all elements in the entire vector instead of dividing the combined feature vector m×n into sub-regions.

[0061] Hybrid pooling can be to divide the combined feature vector m×n into several sub-regions, and randomly select the maximum pooling or average pooling method for each sub-region, thereby preventing overfitting.

[0062] Power average pooling is based on the combination of average pooling and maximum pooling. It uses a learning parameter to determine the relative importance of these two pooling methods. When the learning parameter has different values, different pooling methods are selected.

[0063] Step S104, concatenate multiple intermediate representation vectors to obtain a final representation vector representing the case and the current seat.

[0064] In the embodiment of the present application, step S104 can be implemented in the feature aggregation and prediction layer of the neural network model.

[0065] Exemplarily, assuming that step S103 adopts six different pooling methods for the combined feature vector, then these six pooling methods can correspondingly obtain six intermediate representation vectors, which are respectively marked as: P1, P2, P3, P4, P5, and P6.

[0066] Then, in step S104, the feature aggregation and prediction layer can concatenate the six intermediate representation vectors to obtain a final representation vector Z = [P1, P2, P3, P4, P5, P6].

[0067] Step S105: Output the score of the final representation vector through the activation function. The score is the estimated probability of completing the case target corresponding to the case and the current seat.

[0068] In the embodiment of the present application, step S105 can be implemented in the feature aggregation and prediction layer of the neural network model.

[0069] In the embodiment of the present application, the activation function is used to convert the final representation vector into a score between [0, 1] and output it. The score is an estimated probability that the case goal can be completed under the current allocation relationship between cases and seats.

[0070] Among them, the activation function can be a sigmoid function, a softmax function, etc., which is not limited in the embodiment of the present application.

[0071] It should be noted here that in order to obtain the probability estimate of any case in the case set and any seat combination in the seat set, the embodiment of the present application can input the seat portrait information, called portrait information, case information, called historical call content, real-time status of the agent and other information of each case and seat combination into the neural network model respectively. In this way, the neural network model can generate the probability estimate of the completion of the case goal corresponding to each seat for each case.

[0072] To facilitate the description of the probability estimates of different combinations below, pij is used to represent the probability estimate corresponding to any case i and any seat j, where i represents any i-th case in the case set, and j represents any j-th seat in the seat set.

[0073] Step S106, assigning seats to the multiple cases according to the probability estimation values ​​corresponding to the multiple cases and the multiple seats.

[0074] In the specific implementation, each case in the case set and each seat in the seat set can be used as a node to construct a network flow graph N from the case set to the seat set, and the minimum cost maximum flow of the network flow graph N is solved to obtain the optimal solution for the distribution relationship between multiple seats and multiple cases.

[0075] Figure 6 is a schematic diagram of a network flow graph for solving the minimum cost maximum flow problem in an embodiment of the present application. In one implementation, Figure 6 As shown, the network flow graph N can be defined as N==(V, E, c, w, s, t). Where: V represents the set of nodes, including the nodes of all cases in the case set and the nodes of all seats in the seat set. If it is specific to the collection scenario, the set of nodes includes all nodes A, B, C... in the case set and all nodes 1, 2, 3, 4, 5... in the collection agent set; E represents the set of edges, including the edges between the nodes of all cases in the case set and the nodes of all seats in the seat set, and the edge (vi, vj)∈E between any case i and any seat j; c is the edge capacity between any case i and any seat j, and the edge capacity between any case i and any seat j is defined as 1 here; w is the cost on the edge, and the cost on the edge between any case i and any seat j w(vi, vj)=c-pij, pij is the probability estimate corresponding to any case i and any seat j, s is the origin of the network flow graph, and t is the sink of the network flow graph.

[0076] Next, a maximum flow f is solved for the network flow graph N, so that the total cost of the flow W(f) = ∑w(vi,vj) × f(vi,vj) is the minimum value. Then, the allocation method of cases and seats corresponding to the maximum flow f is the optimal solution for the allocation relationship between multiple seats and multiple cases to be obtained in the embodiment of this application. Then, seats can be allocated to multiple cases according to the optimal solution, so that appropriate cases are allocated to appropriate seats, ensuring the fairness of case allocation, which is conducive to improving the completion rate of case goals.

[0077] It can be seen from the above technical solutions that the embodiment of the present application provides a method for allocating outbound business cases. Based on the neural network model, the potential relationship between the combination of outbound business cases and seats and the results of completing the case is deeply mined, and the deep neural network model is used to predict the probability of all case allocation combinations being able to complete the case goals. Based on the prediction results, a network flow graph is constructed and the optimal allocation combination between cases and seats is found, thereby allocating suitable cases to suitable seats, which not only ensures the fairness of case allocation, but also improves the completion rate of case goals, and is beneficial to improving the work enthusiasm of the agents. It can be seen that the technical solution of the embodiment of the present application solves the problem that cases and seats in current outbound business scenarios such as collection or marketing cannot be optimally allocated.

[0078] The various method steps disclosed in the embodiments of the present application can be implemented by computer software, hardware, or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0079] Figure 7 Schematic diagram of the structure of an outbound call service case allocation device provided in an embodiment of the present application. Figure 7 As shown, the device includes a software module for implementing the outbound call service case allocation device provided in the embodiment of the present application, specifically including: an input module 401, an encoding module 402, a combination and interaction module 403 and a feature aggregation and prediction module 404, which can be used to generate a probability estimate of the completion of the case target corresponding to each seat for each case, wherein:

[0080] Input module 401 is used to obtain the current agent portrait information, case called portrait information, case information and called historical call content.

[0081] The encoding module 402 is used to generate the current agent portrait information, the case called portrait information, the first convolution feature vector corresponding to the case information, and the second convolution feature vector corresponding to the called party's historical call content.

[0082] The combination and interaction module 403 is used to perform a cross product on the first convolution feature vector and the second convolution feature vector to obtain a combined feature vector.

[0083] The combination and interaction module 403 is further used to perform pooling processing on the combined feature vector using a plurality of different pooling methods to obtain a plurality of intermediate representation vectors of the combined feature vector.

[0084] The feature aggregation and prediction module 404 is used to concatenate multiple intermediate representation vectors to obtain a final representation vector representing the case and the current seat.

[0085] The feature aggregation and prediction module 404 is also used to output the score of the final representation vector through an activation function, and the score is an estimated probability of completion of the case target corresponding to the case and the current agent.

[0086] The case allocation device also includes:

[0087] The solution module 405 is used to allocate seats for the multiple cases according to the probability estimation values ​​corresponding to the multiple cases and the multiple seats.

[0088] Figure 8Schematic diagram of the structure of an outbound call service case allocation device provided in an embodiment of the present application. Figure 8 As shown, the device includes a hardware module for implementing the outbound call service case allocation device provided in the embodiment of the present application, including: a processor 501 and a memory 502, the memory 502 includes a program instruction 503, when the program instruction 503 is executed by the processor 501, the outbound call service case allocation device performs the following method steps:

[0089] For each case, the probability estimate of the completion of the case target corresponding to each agent is generated respectively. The probability estimate of the completion of the case target corresponding to each agent includes:

[0090] Generate the current agent portrait information, the case called portrait information, the first convolution feature vector corresponding to the case information, and generate the second convolution feature vector corresponding to the called party's historical call content;

[0091] Cross-multiply the first convolution feature vector and the second convolution feature vector to obtain a combined feature vector;

[0092] Using a variety of different pooling methods to perform pooling processing on the combined feature vector to obtain multiple intermediate representation vectors of the combined feature vector;

[0093] Multiple intermediate representation vectors are concatenated to obtain the final representation vector representing the case and the current seat.

[0094] The activation function outputs the score of the final representation vector, which is the estimated probability of completing the case goal corresponding to the case and the current seat.

[0095] Seat allocation for multiple cases is performed based on the probability estimates corresponding to the multiple cases and the multiple seats.

[0096] It can be seen from the above technical scheme that the embodiment of the present application provides an outbound business case allocation device, which deeply mines the potential relationship between the combination of outbound business cases and seats and the completion results of the case based on the neural network model, and uses the deep neural network model to predict the probability of all case allocation combinations to complete the case goals, and constructs a network flow graph based on the prediction results to find the optimal allocation combination between cases and seats, thereby realizing the allocation of suitable cases to suitable seats, which not only ensures the fairness of case allocation, but also improves the completion rate of case goals, and is also conducive to improving the work enthusiasm of the seats. It can be seen that the technical solution of the embodiment of the present application solves the problem that cases and seats in current outbound business scenarios such as collection or marketing cannot be optimally allocated.

[0097] It is easy to understand that those skilled in the art can combine, split, reorganize, etc. the embodiments of the present application to obtain other embodiments based on the several embodiments provided in the present application, and these embodiments do not exceed the protection scope of the present application.

[0098] The above specific implementation methods further explain in detail the purpose, technical solutions and beneficial effects of the embodiments of the present application. It should be understood that the above are only specific implementation methods of the embodiments of the present application and are not used to limit the protection scope of the embodiments of the present application. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solutions of the embodiments of the present application should be included in the protection scope of the embodiments of the present application.

Claims

1. A method for allocating outbound call service cases, characterized in that: include: For each case, a probability estimate of the completion of the case target corresponding to each agent is generated respectively. The generating of the probability estimate of the completion of the case target corresponding to each agent includes: Generate the current agent portrait information, the case called person portrait information, the first convolution feature vector corresponding to the case information, and generate the second convolution feature vector corresponding to the called person's historical call content; the agent portrait information is information related to the agent itself in the outbound call business scenario, the called person portrait information is information related to the called person itself in the outbound call business scenario, the case information is information related to the case itself in the outbound call business scenario, and the called person's historical call content is the call content generated during the call between the outbound call robot and the called person before the case is assigned to a certain agent, or the call content between the called person and the agent before the call; Cross-multiplying the first convolution feature vector and the second convolution feature vector to obtain a combined feature vector; Using a plurality of different pooling methods to perform pooling processing on the combined feature vector to obtain a plurality of intermediate representation vectors of the combined feature vector; splicing the multiple intermediate representation vectors to obtain a final representation vector representing the case and the current seat; Outputting the score of the final representation vector through an activation function, wherein the score is an estimated probability value of the completion of the case target corresponding to the case and the current seat; Allocate seats for multiple cases based on the estimated probability values ​​corresponding to multiple cases and multiple seats; The allocating seats for the multiple cases according to the probability estimation values ​​corresponding to the multiple cases and the multiple seats includes: Taking each case in the case set and each seat in the seat set as a node, a network flow graph N between the case set and the seat set is constructed, and the minimum cost maximum flow of the network flow graph N is solved to obtain the optimal solution of the allocation relationship between the multiple seats and the multiple cases; The network flow graph N = (V, E, c, w, s, t), wherein V represents a set of nodes, E represents a set of edges, the edge (vi, vj)∈E between any case i and any seat j, c is the edge capacity, w is the cost on the edge, the cost on the edge between any case i and any seat j w(vi, vj)=c-pij, pij is the probability estimate corresponding to any case i and any seat j, s is the origin of the network flow graph, and t is the sink of the network flow graph.

2. The method according to claim 1, characterized in that The generating of the first convolution feature vector corresponding to the current agent portrait information, the case called portrait information, and the case information includes: Encoding the current agent portrait information, the case called portrait information, and the structured text of the case information to obtain a first text feature vector; A convolutional neural network is used to extract convolutional features from the first text feature vector to obtain the first convolutional feature vector.

3. The method according to claim 1, characterized in that The generating of the second convolution feature vector corresponding to the historical call content of the called party includes: Convert the called party's historical call content into dialogue text information using speech synthesis technology; Encoding the conversation text information to obtain a corresponding second text feature vector; The second text feature vector is processed using a convolutional neural network to obtain the second convolution feature vector.

4. The method according to claim 3, characterized in that The using a convolutional neural network to process the second text feature vector to obtain the second convolution feature vector includes: Use multiple convolution kernels of different sizes to extract convolution features from the second text feature vector respectively, and concatenate the extracted convolution features to obtain the second convolution feature vector.

5. The method according to claim 1, characterized in that The pooling methods include at least two of the following: maximum pooling, average pooling, global average pooling, mixed pooling, random pooling, and power average pooling.

6. The method according to claim 1, characterized in that The activation function is a sigmoid activation function.

7. The method according to claim 1, characterized in that When the outbound call service case is a collection case, the agent portrait information is the collection agent portrait information, the called party portrait information is the collection object portrait information, and the case information at least includes the case amount and the number of overdue days.

8. According to the method described in any one of claims 1-7, when outbound call business cases are allocated in real time, the information used to generate the first convolution feature vector also includes the real-time status of the agent.

9. An outbound call service case allocation device, characterized in that: include: The input module, encoding module, combination and interaction module, and feature aggregation and prediction module are used to generate the probability estimate of the completion of the case goal corresponding to each seat for each case, where: The input module is used to obtain the current agent portrait information, the case called portrait information, the case information and the called party's historical call content; The encoding module is used to generate the current seat portrait information, the case called portrait information, the first convolution feature vector corresponding to the case information, and generate the second convolution feature vector corresponding to the called party's historical call content; the seat portrait information is the information related to the seat itself in the outbound call business scenario, the called party portrait information is the information related to the called party itself in the outbound call business scenario, the case information is the information related to the case itself in the outbound call business scenario, and the called party's historical call content is the call content generated during the call between the outbound call robot and the called party before the case is assigned to a certain seat, or the call content between the called party and the seat before the call; The combination and interaction module is used to perform a cross product of the first convolution feature vector and the second convolution feature vector to obtain a combined feature vector; The combination and interaction module is further used to perform pooling processing on the combined feature vector using a plurality of different pooling methods to obtain a plurality of intermediate representation vectors of the combined feature vector; The feature aggregation and prediction module is used to concatenate multiple intermediate representation vectors to obtain a final representation vector representing the case and the current seat; The feature aggregation and prediction module is further used to output the score of the final representation vector through an activation function, wherein the score is an estimated probability value of completion of the case target corresponding to the case and the current seat; The case allocation device also includes: A solution module, used for allocating seats for multiple cases according to probability estimation values ​​corresponding to multiple cases and multiple seats; The solution module is specifically used for: Taking each case in the case set and each seat in the seat set as a node, a network flow graph N between the case set and the seat set is constructed, and the minimum cost maximum flow of the network flow graph N is solved to obtain the optimal solution of the allocation relationship between the multiple seats and the multiple cases; The network flow graph N = (V, E, c, w, s, t), wherein V represents a set of nodes, E represents a set of edges, the edge (vi, vj)∈E between any case i and any seat j, c is the edge capacity, w is the cost on the edge, the cost on the edge between any case i and any seat j w(vi, vj)=c-pij, pij is the probability estimate corresponding to any case i and any seat j, s is the origin of the network flow graph, and t is the sink of the network flow graph.

Images