Method and apparatus for predicting voice interaction path, and storage medium

Abstract

The application discloses a voice interaction path prediction method and device and a storage medium. The method comprises the following steps: in the case that a target voice call initiated by a target user number is received, a resource operation feature vector corresponding to a resource operation feature set associated with the target user number is acquired; a node feature vector corresponding to a pre-path node completed by the target user number in the target voice call and an end point feature vector corresponding to all path end points associated with the target voice call are determined; and a target path node is predicted based on a target feature vector to be predicted determined by the resource operation feature vector, the node feature vector and the end point feature vector, the target path node being a next path node after a current completed voice interaction sub-service. The application solves the technical problem of low prediction result accuracy caused by the need to constantly manually adjust key parameters in the voice interaction path prediction method provided in the related art.

Description

TECHNICAL FIELD

[0001] The present application relates to the computer field, in particular, to a voice interaction path prediction method and device and a storage medium. BACKGROUND

[0002] In order to improve the quality of consulting services, intelligent voice interaction systems are introduced in many telephone customer service businesses to automatically interact with customers. A common interaction scenario is that when a customer calls the background, the intelligent voice interaction system can ask the customer to confirm whether to handle a certain type of business. If it is to handle a target type of business configured with an interactive voice response (IVR) menu, the customer will be navigated to the corresponding IVR process to achieve the purpose of distributing the IVR menu and the human operator. For example, in the context of a commercial bank, when the customer selects a credit card service category, the system will further ask questions about the service items it contains, such as "Do you want to check your credit card bill?"

[0003] That is, in the process of interaction between the intelligent voice interaction system and the customer, the voice interaction path composed of the IVR menu frequently visited by the customer needs to be excavated in advance, so that the system can determine the candidate options for the next service node in advance when the customer selects a non-final service node, to provide the customer with a reference for answering.

[0004] However, the current voice interaction path prediction method provided in the related art is usually strongly dependent on the key parameters in the existing frequent path excavation algorithm. These key parameters are mainly selected manually by business experience, and are also affected by the external business environment and need to be adjusted manually from time to time. In other words, since the existing voice interaction path prediction method needs to manually adjust the key parameters based on business experience, it is difficult to ensure the accuracy of the prediction results.

[0005] To address the above problems, no effective solutions have been proposed. SUMMARY

[0006] The embodiments of the present application provide a voice interaction path prediction method and device and a storage medium to at least solve the technical problem of low accuracy of prediction results caused by the need to constantly manually adjust the key parameters in the voice interaction path prediction method provided in the related art.

[0007] According to an aspect of the embodiments of the present application, a method for predicting a voice interaction path is provided. The method comprises: in response to receiving a target voice call initiated by a target user number, obtaining a resource operation feature vector corresponding to a resource operation feature set associated with the target user number, wherein the resource operation feature set comprises resource operation features of the target user number obtained under each channel in a resource management platform; determining a node feature vector corresponding to a preceding path node completed by the target user number in the target voice call, and an end point feature vector corresponding to all path end points associated with the target voice call, wherein the preceding path node is a path node corresponding to a voice interaction sub-service currently completed by the target user number; and predicting a target path node based on a target feature vector to be predicted determined based on the resource operation feature vector, the node feature vector and the end point feature vector, wherein the target path node is a path node corresponding to a target voice interaction sub-service to be executed next after the voice interaction sub-service currently completed by the target user number.

[0008] Optionally, in the embodiments, the predicting the target path node based on the target feature vector to be predicted determined based on the resource operation feature vector, the node feature vector and the end point feature vector comprises: inputting the target feature vector determined into a path prediction model, wherein the path prediction model is used to calculate a prediction probability of each candidate path node being the target path node; and determining the target path node from the candidate path nodes according to a sorting result of the prediction probability.

[0009] Optionally, in the embodiments, before the inputting the target feature vector determined into the path prediction model, the method further comprises: performing compression transformation on the resource operation feature vector to obtain a first intermediate vector; performing adjustment and update on the end point feature vector based on the node feature vector and a node mapping weight to obtain a second intermediate vector; and splicing the first intermediate vector and the second intermediate vector to obtain the target feature vector.

[0010] Optionally, in the embodiment, the obtaining of the resource operation feature vector corresponding to the resource operation feature set associated with the target user number comprises: obtaining account attribute information of a target user resource account registered by the target user number in the resource management platform, and historical resource operation records of each channel associated with the target user resource account; obtaining historical call records initiated by all user numbers before the target voice call; extracting channel features corresponding to the target user resource account from the account attribute information and the historical resource operation records and the historical call records, and extracting path features from the historical call records to obtain the resource operation feature set; and encoding the resource operation feature set by a double-layer random forest model to obtain the resource operation feature vector.

[0011] Optionally, in the embodiment, before the obtaining of the resource operation feature vector corresponding to the resource operation feature set associated with the target user number, the method further comprises: obtaining sample data of sample voice calls initiated by a plurality of sample user numbers; determining sample resource operation feature vectors associated with the sample user numbers; encoding non-path end nodes in all path nodes of the sample voice calls to obtain sample node feature vectors, and encoding all path ends of the sample voice calls to obtain sample end feature vectors; and training an initialized path prediction model based on sample feature vectors determined based on the sample resource operation feature vectors, the sample node feature vectors, and the sample end feature vectors, until a convergence condition is reached.

[0012] Optionally, in the embodiment, the training of the initialized path prediction model based on the sample feature vectors determined based on the sample resource operation feature vectors, the sample node feature vectors, and the sample end feature vectors comprises: performing compression transformation on the sample resource operation feature vectors to obtain a first sample intermediate vector; updating the sample end feature vectors based on the sample node feature vectors and a training-adjusted node mapping weight to obtain a second sample intermediate vector, wherein the node mapping weight is a mapping weight between the non-path end nodes and the path ends; splicing the first sample intermediate vector and the second sample intermediate vector to obtain the sample feature vectors; and training the initialized path prediction model by using the sample feature vectors.

[0013] Optionally, in the embodiment, the training of the initialized path prediction model based on the sample feature vector determined based on the sample resource operation feature vector, the sample node feature vector and the sample terminal feature vector includes: in a case where a training output result corresponding to the sample feature vector and a label matched with the sample feature vector do not match, a training error corresponding to the training output result is back propagated to adjust a training weight in the path prediction model.

[0014] According to another aspect of the embodiment of the present application, a voice interaction path prediction device is further provided, which includes: an acquisition unit configured to acquire a resource operation feature vector corresponding to a resource operation feature set associated with a target user number in a case where a target voice call initiated by the target user number is received, wherein the resource operation feature set includes resource operation features of the target user number acquired under each channel in a resource management platform; a determination unit configured to determine a node feature vector corresponding to a preceding path node completed by the target user number in the target voice call, and a terminal feature vector corresponding to all path terminals associated with the target voice call, wherein the preceding path node is a path node corresponding to a voice interaction sub-service currently completed by the target user number; and a prediction unit configured to predict a target path node based on a target feature vector determined based on the resource operation feature vector, the node feature vector and the terminal feature vector, wherein the target path node is a path node corresponding to a target voice interaction sub-service to be executed next after the voice interaction sub-service currently completed by the target user number.

[0015] Optionally, in the embodiment, the prediction unit includes: an input module configured to input the determined target feature vector into a path prediction model, wherein the path prediction model is configured to calculate a prediction probability of each candidate path node being the target path node; and a prediction module configured to determine the target path node from the candidate path nodes according to a sorting result of the prediction probability.

[0016] According to still another aspect of the embodiment of the present application, a computer readable storage medium is further provided, and the computer readable storage medium stores a computer program, wherein the computer program is configured to execute the voice interaction path prediction method when running.

[0017] In the embodiment of the present application, by using the method provided in the embodiment, the resource operation feature vector corresponding to the user number initiating the call is obtained based on the user number, and the target feature vector to be predicted is determined based on the resource operation feature vector, the node feature vector of the pre-path node completed by the user number, and the terminal feature vector of all path terminals, and the target path node to be executed next by the user number is predicted through the path prediction model obtained by automatic training, so as to intelligently predict the target path node to be executed next by the user number based on the channel feature associated with the user number, accurately mine the voice interaction path that the user number may use, and improve the accuracy of voice interaction. Further, the technical problem of low prediction result accuracy caused by the need to manually adjust key parameters in related technologies is overcome. BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings used to provide further understanding of the present application and constitute a part of the present application, and the illustrative embodiments of the present application and the description thereof are used to explain the present application, and do not constitute improper limitations on the present application. In the drawings:

[0019] Figure 1 is a flowchart of an optional voice interaction path prediction method according to an embodiment of the present application;

[0020] Figure 2 is a schematic diagram of an optional voice interaction path prediction method according to an embodiment of the present application;

[0021] Figure 3 is a structural schematic diagram of an optional voice interaction path prediction device according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0023] It should be noted that the terms "first", "second", and the like in the description and in the claims of the present application and in the above description of the drawings merely mean different instances of similar objects, and do not necessarily imply a specific order or sequence. It should be understood that the data thus used in the description interchangeably can be interchanged where appropriate, so that the embodiments of the present application described herein can be implemented in other than the order shown or described herein. In addition, the terms "comprise" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that includes a list of steps or units as an example is not necessarily limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to such processes, methods, products, or apparatuses.

[0024] According to an aspect of the embodiments of the present application, a method for predicting a voice interaction path is provided. The flow of the method can be as shown in Figure 1

[0025] S102, in the case of receiving a target voice call initiated by a target user number, obtaining a resource operation feature vector corresponding to a resource operation feature set associated with the target user number, wherein the resource operation feature set includes resource operation features of the target user number obtained under each channel in a resource management platform;

[0026] S104, determining a node feature vector corresponding to a preceding path node completed by the target user number in the target voice call, and an end point feature vector corresponding to all path end points associated with the target voice call, wherein the preceding path node is a path node corresponding to a voice interaction sub-service that has been completed by the target user number;

[0027] S106, predicting a target path node based on a target feature vector to be predicted determined by the resource operation feature vector, the node feature vector, and the end point feature vector, wherein the target path node is a path node corresponding to a target voice interaction sub-service to be executed next after the voice interaction sub-service that has been completed by the target user number.

[0028] ​Optionally, in the embodiment, the prediction method of the voice interaction path can be applied to various intelligent voice interaction scenarios that need to provide voice interaction consultation services, such as bank voice interaction services, voice interaction services of mobile call services, and the like, which are intelligent automatic question and answer processes completed by a computer system and a user. By using the method provided in the embodiment, based on the user number initiating the call, the resource operation feature vector corresponding to the user number is obtained, and based on the resource operation feature vector, the node feature vector of the pre-path node completed by the user number and the end feature vector of all path ends, the target feature vector to be predicted is determined, and the target path node to be executed next by the user number is predicted through the path prediction model obtained by automatic training, so as to intelligently predict the target path node to be executed next by the user number based on the channel feature associated with the user number, accurately mine the voice interaction path that can be used by the user number, and improve the accuracy of voice interaction. Further, the technical problem of low prediction result accuracy caused by the need to manually adjust key parameters in the related art is overcome. Further, in the case of improving the accuracy of the voice interaction path, the user will not need to repeatedly perform a back operation, and the service requested to be completed by the current voice call can be more efficiently completed.

[0029] It should be noted that the prediction algorithm of the voice interaction path generally provided in the related art mainly includes an Apriori algorithm and an FP-Tree algorithm. The Apriori algorithm can but is not limited to mine a frequent path generated in a voice interaction process through three key parameters of minimum support, minimum confidence, and maximum item set number. The support degree here refers to the frequency or probability of occurrence of a path. The confidence degree here refers to the probability of occurrence of a subsequent node when a preceding node of the path occurs. Only when the length of a path is less than or equal to the maximum item set number, and the support degree and the confidence degree are greater than the minimum support degree and the minimum confidence degree, the path is determined as a frequent path. FP-Tree: As an algorithm for mining frequent item sets, the Apriori algorithm needs to scan data multiple times, and I / O is a big bottleneck. In order to solve this problem, the FP Tree algorithm (also known as the FP Growth algorithm) uses some techniques, and only needs to scan the data set twice regardless of the amount of data, thereby improving the efficiency of algorithm operation. The algorithm includes: 1) establishing an item head table and establishing a sorted (deleting non-frequent 1-item set) data set; 2) establishing an FP tree: reading the sorted data set and inserting the FP tree; 3) association rule mining: finding the conditional pattern base corresponding to the item head table item from the bottom item of the item head table upward. The frequent item set of the item head table item is recursively mined from the conditional pattern base.

[0030] Optionally, in the embodiment, the resource operation feature set can include, but is not limited to, the account attribute information of the target user resource account registered by the target user number of the target voice call, such as the resource quantity and resource distribution associated with the target user resource account under various channels, the education information and work information of the target user account, and the operation activity of the target user account. The resource operation feature set can also include, but is not limited to, the path information constituting the frequent path in all call processes. By using the channel features corresponding to the frequent path and the channel information for joint modeling to complete the training of the path prediction model, the frequent path weight concept is given, the intelligent path selection effect is achieved, and the channel feature influence path selection ability is also given. Different characteristics but the same pre-inbound path of customers will get different recommendations. It is no longer strongly dependent on the business experience and space-time attributes in the related art, so that the prediction result is more accurate and objective.

[0031] Through the embodiments provided in the present application, based on the target user number initiating the call, the resource operation feature vector corresponding to the target user number is obtained, and based on the resource operation feature vector, the node feature vector of the pre-path node completed by the target user number and the end feature vector of all path endpoints are determined to determine the target feature vector to be predicted, so as to predict the target path node to be executed next by the user number. Thus, based on the channel features associated with the user number, the target path node to be executed next by the user number is intelligently predicted, the voice interaction path possibly used by the user number is accurately mined, and the accuracy of voice interaction is improved. Further, the technical problem of low prediction result accuracy caused by the need for constant manual adjustment of key parameters in the related art is overcome.

[0032] As an optional solution, based on the resource operation feature vector, the node feature vector and the end feature vector, the target path node to be predicted is predicted, including:

[0033] S1, input the determined target feature vector into the path prediction model, wherein the path prediction model is used to calculate the prediction probability of each candidate path node being the target path node;

[0034] S2, according to the sorting result of the prediction probability, determine the target path node from the candidate path nodes.

[0035] Optionally, in this embodiment, the path prediction model described above can be, but is not limited to, a model based on a softmax regression model for predicting the probability of the target path node. Wherein, the model can be, but is not limited to, trained by using a BP algorithm, and deployed after the training is completed. Wherein, the BP algorithm is a learning process composed of two processes of forward propagation of signals and backward propagation of errors. Wherein, when the forward propagation, the input sample is transmitted by the input layer, processed layer by layer through each hidden layer, and transmitted to the output layer. If the actual output of the output layer is different from the expected output, the error back propagation result is entered. Back propagation is to transmit the output error to the input layer through the hidden layer in a certain form, so that each layer obtains an error signal, and corrects its own training weight based on the error signal.

[0036] It should be noted that in this embodiment, the output result of the path prediction model contains the feasibility prediction probability of each path node being the target path node.

[0037] Through the embodiments provided in this application, the target path node matched with the target feature vector is predicted by the path prediction model. Here, the path prediction model gives different weights to different voice interaction paths in the automatic training process, so as to realize intelligent sorting of paths with the same starting point, so as to more accurately and efficiently predict the voice interaction path that the target user number may select in this voice call, and the next target path node after the current completed pre-path node in the voice interaction path.

[0038] As an optional solution, before inputting the determined target feature vector into the path prediction model, the following steps are further included:

[0039] S1, compressing and transforming the resource operation feature vector to obtain a first intermediate vector;

[0040] S2, adjusting and updating the terminal feature vector based on the node feature vector and the node mapping weight to obtain a second intermediate vector;

[0041] S3, splicing the first intermediate vector and the second intermediate vector to obtain the target feature vector.

[0042] Optionally, in this embodiment, the compression transformation described above can be, but is not limited to, a deep neural network (Deep Neural Networks, DNN) and batch normalization (Batch Normalization).

[0043] Optionally, in the embodiment, the node mapping weight can be but is not limited to be used for connecting the node of the non-path end point and the path end point. The node mapping weight can be but is not limited to be initialized in the form of 0-1 according to the actually existing path start point-end point correspondence. In addition, the node mapping weight can be but is not limited to be configured with a penalty factor, including an upward penalty factor a and a downward penalty factor β, to ensure that the integer value is not deviated too much in the subsequent training process.

[0044] Through the embodiments provided in the application, the first intermediate vector carrying the channel characteristics and the second intermediate vector carrying the path characteristics are spliced to obtain a target feature vector used for outputting the path prediction model for prediction, so as to combine the channel data domain in the voice interaction process with the traditional path data, thereby making the predicted path result more diversified.

[0045] As an optional solution, obtaining the resource operation feature vector corresponding to the resource operation feature set associated with the target user number comprises:

[0046] S1, obtaining account attribute information of a target user resource account registered by a target user number in a resource management platform, and historical resource operation records of the target user resource account under each channel;

[0047] S2, obtaining historical call records initiated by all user numbers before a target voice call;

[0048] S3, extracting channel characteristics corresponding to the target user resource account from the account attribute information and the historical resource operation records and the historical call records to obtain a resource operation feature set;

[0049] S4, encoding the resource operation feature set by a double-layer random forest model to obtain a resource operation feature vector.

[0050] It should be noted that the path characteristics can be but are not limited to be obtained by a traditional method, such as an Apriori algorithm and an FP-Tree algorithm. The channel characteristics can be but are not limited to be attribute characteristics and operation characteristics associated with each channel of the platform by the user number initiating the voice call.

[0051] Through the embodiments provided in the application, the channel characteristics are combined with the frequent path to make the voice interaction path have more diversified changes, thereby achieving the effect of intelligently recommending more accurate options for the user in the voice interaction question and answer process.

[0052] As an optional solution, before obtaining the resource operation feature vector corresponding to the resource operation feature set associated with the target user number, the method further comprises:

[0053] S1, obtaining sample data of sample voice calls initiated by a plurality of sample user numbers;

[0054] S2, determining a sample resource operation feature vector associated with the sample user number;

[0055] S3, encoding non-path end nodes in all path nodes of the sample voice calls to obtain a sample node feature vector, and encoding all path ends of the sample voice calls to obtain a sample end feature vector;

[0056] S4, training the initialized path prediction model based on the sample feature vector determined by the sample resource operation feature vector, the sample node feature vector and the sample end feature vector, until a convergence condition is reached.

[0057] It should be noted that the sample end feature vector is the initial value of the vector, which will be continuously adjusted and updated according to the obtained node feature vector and the corresponding weight in the subsequent training process.

[0058] For example, the set of all path ends is encoded into a vector C3 as the initialization C3. Assuming that there are 1000 different incoming paths in history, of which there may be 300 different incoming ends. Then C3 is a 300-dimensional vector. Further, in the actual incoming process of the customer, according to the value of C2 that has occurred and the value of the weight matrix W, the value of each dimension of the C3 vector in 300 dimensions can be calculated.

[0059] In addition, the initialization of the above vector C2 is the same. Assuming that 1000 different incoming paths in history can find out 2000 sub-paths in total, then C2 is a 2000-dimensional vector. In the actual incoming process of the customer, since a certain sub-path is triggered, the value of the 2000-dimensional vector is 1 only at that node, and the rest are 0, which is a one-hot vector.

[0060] Optionally, in the embodiment, S4 training the initialized path prediction model based on the sample feature vector determined by the sample resource operation feature vector, the sample node feature vector and the sample end feature vector includes:

[0061] S41, compressing and transforming the sample resource operation feature vector to obtain a first sample intermediate vector;

[0062] S42, adjusting and updating the sample end feature vector based on the sample node feature vector and the training-adjusted node mapping weight to obtain a second sample intermediate vector, wherein the node mapping weight is the mapping weight between the non-path end node and the path end;

[0063] S43, splicing the first sample intermediate vector and the second sample intermediate vector to obtain a sample feature vector;

[0064] S44, training the initialized path prediction model by using the sample feature vector.

[0065] Optionally, in the embodiment, training the initialized path prediction model based on the sample feature vector determined based on the sample resource operation feature vector, the sample node feature vector and the sample terminal feature vector comprises: in the case that a training output result corresponding to the sample feature vector does not match a label matched with the sample feature vector, reversely propagating a training error corresponding to the training output result to adjust a training weight in the path prediction model.

[0066] Specifically, the embodiment is described in combination with the examples shown in the following table: Figure 2 It is assumed that a target voice call initiated by a target user number is a consultation request call under a bank business system.

[0067] Sample data of a sample voice call initiated by a sample user number is obtained, wherein sample user account financial transactions and behavior features of the sample user number registered in the bank under each channel are introduced, a double-layer machine forest model is used, feature importance is calculated by OOB, features available for the bank customer service channel are screened, and the features are encoded into a resource operation feature vector C1.

[0068] The resource operation feature vector C1 is compressed and transformed by DNN+Batch Normalization to obtain an intermediate vector C4.

[0069] All paths and sub-paths (nodes other than path terminals) of the sample user number call process are encoded into a node feature vector C2 corresponding to a pre-path node, and a set of all path terminals is encoded into a terminal feature vector C3.

[0070] Then, a weight matrix W between C2 and C3 is initialized. The weight matrix W is initialized in the form of 0 / 1 according to the actual existing path start point-terminal point corresponding relationship, and an upward penalty factor a and a downward penalty factor b are set to ensure that the integer value is not deviated too much in the subsequent training process.

[0071] After the terminal feature vector C3 adjusted by the weight matrix and the intermediate vector C4 are spliced, a softmax regression model is established to output a probability, and a BP algorithm is used for training until the output result reaches a convergence condition.

[0072] It should be noted that in actual use, after each node is selected, the above path prediction model can be called again to predict a target path node corresponding to a sub-business to be executed next time to complete the current demand of the user.

[0073] In addition, in the embodiment, the model parameters of the path prediction model are optimized according to the results of the voice call generated by the user number in the real scene every day. Since a new call may bring a new voice interaction path, in the embodiment, the vector C2 and the vector C3 can be updated first, but are not limited to this. The W weight matrix is also correspondingly expanded in dimension, such as inheriting the original weight coefficient first, then initializing the weight coefficient between all new nodes in the form of 0 / 1, and finally updating the parameters of the entire model using the BP algorithm.

[0074] The above-mentioned process is an example, and the embodiment is not limited in this regard.

[0075] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited to the action sequence described, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0076] According to another aspect of the embodiment of the present application, a voice interaction path prediction device for implementing the voice interaction path prediction method is also provided. As shown in the figure, the device comprises: Figure 3

[0077] The acquisition unit 302 is configured to, in the case that a target voice call initiated by a target user number is received, acquire a resource operation feature vector corresponding to a resource operation feature set associated with the target user number, wherein the resource operation feature set comprises resource operation features of the target user number acquired under each channel in the resource management platform;

[0078] The determination unit 304 is configured to determine a node feature vector corresponding to a preceding path node completed by the target user number in the target voice call, and an end point feature vector corresponding to all path end points associated with the target voice call, wherein the preceding path node is a path node corresponding to a voice interaction sub-service that has been completed by the target user number.

[0079] The prediction unit 306 is configured to predict a target path node based on a target feature vector to be predicted determined by the resource operation feature vector, the node feature vector and the end point feature vector, wherein the target path node is a path node corresponding to a target voice interaction sub-service to be executed next after the voice interaction sub-service that has been completed by the target user number.

[0080] Optionally, in the embodiment, the prediction unit 306 comprises:​

[0081] inputting the determined target feature vector into a path prediction model, wherein the path prediction model is configured to calculate a prediction probability of each candidate path node being the target path node;

[0082] predicting, by the prediction module, the target path node from the candidate path nodes according to the ranking result of the prediction probabilities.

[0083] Optionally, in the embodiment, the prediction unit 306 further includes:

[0084] compressing the resource operation feature vector to obtain a first intermediate vector before inputting the determined target feature vector into the path prediction model;

[0085] adjusting the end point feature vector based on the node feature vector and the node mapping weight to obtain a second intermediate vector;

[0086] splicing the first intermediate vector and the second intermediate vector to obtain the target feature vector.

[0087] Optionally, in the embodiment, the obtaining unit 302 includes:

[0088] a first obtaining module configured to obtain account attribute information of a target user resource account registered by a target user number in a resource management platform, and historical resource operation records of each channel associated with the target user resource account;

[0089] a second obtaining module configured to obtain historical call records initiated by all user numbers before the target voice call;

[0090] an extracting module configured to extract channel features corresponding to the target user resource account from the account attribute information and the historical resource operation records and the historical call records to obtain a resource operation feature set;

[0091] an encoding module configured to encode the resource operation feature set by a double-layer random forest model to obtain a resource operation feature vector.

[0092] Optionally, in the embodiment, the method further includes:

[0093] a training obtaining unit configured to obtain sample data of sample voice calls initiated by a plurality of sample user numbers before obtaining a resource operation feature vector corresponding to a resource operation feature set associated with the target user number;

[0094] a training determining unit configured to determine a sample resource operation feature vector associated with the sample user number;

[0095] The training coding unit is configured to code all non-path end nodes in all path nodes of the sample voice call to obtain a sample node feature vector, and code all path ends of the sample voice call to obtain a sample end feature vector;

[0096] The training unit is configured to train the initialized path prediction model based on the sample feature vector determined based on the sample resource operation feature vector, the sample node feature vector, and the sample end feature vector, until a convergence condition is reached.

[0097] Optionally, in the embodiment, the training unit comprises:

[0098] The training compression module is configured to perform compression transformation on the sample resource operation feature vector to obtain a first sample intermediate vector.

[0099] The training adjustment module is configured to adjust and update the sample end feature vector based on the sample node feature vector and the adjusted node mapping weight to obtain a second sample intermediate vector, wherein the node mapping weight is a mapping weight between the non-path end node and the path end.

[0100] The training splicing module is configured to splice the first sample intermediate vector and the second sample intermediate vector to obtain the sample feature vector.

[0101] The training module is configured to train the initialized path prediction model by using the sample feature vector.

[0102] Optionally, in the embodiment, the training unit comprises:

[0103] The error adjustment module is configured to, in a case where a training output result corresponding to the sample feature vector does not match a label matched with the sample feature vector, perform back propagation of a training error corresponding to the training output result to adjust a training weight in the path prediction model.

[0104] It should be noted that, in the embodiment, the embodiments to be implemented by the above-mentioned various unit modules can refer to the above-mentioned various method embodiments, which will not be described herein.

[0105] According to an aspect of the present application, a computer readable storage medium is provided, and a processor of a computer device reads computer instructions from the computer readable storage medium. The processor executes the computer instructions, so that the computer device executes the voice interaction path prediction method provided in the various optional implementation manners.

[0106] Optionally, in the embodiment, the above-mentioned computer readable storage medium can be configured to store a computer program for executing the following steps:

[0107] S1, in the case of receiving a target voice call initiated by a target user number, obtaining a resource operation feature vector corresponding to a resource operation feature set associated with the target user number, wherein the resource operation feature set comprises resource operation features of the target user number obtained under each channel in a resource management platform;

[0108] S2, determining a node feature vector corresponding to a pre-path node completed by the target user number in the target voice call, and an end point feature vector corresponding to all path end points associated with the target voice call, wherein the pre-path node is a path node corresponding to a voice interaction sub-service currently completed by the target user number;

[0109] S3, predicting a target path node based on a target feature vector to be predicted determined by the resource operation feature vector, the node feature vector and the end point feature vector, wherein the target path node is a path node corresponding to a target voice interaction sub-service to be predicted for the target user number after the voice interaction sub-service currently completed.

[0110] Optionally, in the embodiment, all or part of the steps in the various methods of the above-described embodiments can be completed by a program instructing the hardware related to the terminal device, and the program can be stored in a computer readable storage medium, which can include a flash disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0111] The integrated units in the above-described embodiments, if realized in the form of software function units and sold or used as independent products, can be stored in the above-described computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium, includes a number of instructions to make one or more computer devices (which can be personal computers, servers or network devices, etc.) execute all or part of the steps of the methods described in the various embodiments of the present application.

[0112] In the above-described embodiments of the present application, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0113] In several embodiments provided in the present application, it should be understood that the disclosed client can be implemented in other manners. Of course, the described apparatus embodiments are merely schematic, and the division of units is merely logical function division, and there can be other division manners in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, units or modules, and can be in electrical, mechanical or other forms.

[0114] The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments of the present application.

[0115] In addition, each functional unit in the embodiments of the present application can be integrated in a processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be implemented in the form of hardware, or in the form of a software functional unit.

[0116] The above is merely the preferred embodiments of the present application, and it should be pointed out that, for those skilled in the art, some improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A method of predicting a voice interaction path, characterized by, The method comprises the following steps: acquiring sample data of sample voice calls initiated by a plurality of sample user numbers; determining a sample resource operation feature vector associated with the sample user numbers; encoding non-path end nodes in all path nodes of the sample voice calls to obtain sample node feature vectors, and encoding all path ends of the sample voice calls to obtain sample end feature vectors; training an initialized path prediction model based on a sample feature vector determined by the sample resource operation feature vector, the sample node feature vector and the sample end feature vector until a convergence condition is reached; in the case of receiving a target voice call initiated by a target user number, acquiring a resource operation feature vector corresponding to a resource operation feature set associated with the target user number, wherein the resource operation feature set comprises resource operation features of the target user number acquired under each channel in a resource management platform; determining a node feature vector corresponding to a pre-path node completed by the target user number in the target voice call, and an end feature vector corresponding to all path ends associated with the target voice call, wherein the pre-path node is a path node corresponding to a voice interaction sub-service currently completed by the target user number; inputting a target feature vector to be predicted, which is determined based on the resource operation feature vector, the node feature vector and the end feature vector, into a path prediction model, wherein the path prediction model is used to calculate a prediction probability of each candidate path node being a target path node, the target path node being a path node corresponding to a target voice interaction sub-service to be predicted for the target user number after the currently completed voice interaction sub-service; determining the target path node from the candidate path nodes according to a ranking result of the prediction probability.

2. The method of claim 1, wherein, Before inputting the target feature vector to be predicted, which is determined based on the resource operation feature vector, the node feature vector and the end feature vector, into the path prediction model, the method further comprises the following steps: performing compression transformation on the resource operation feature vector to obtain a first intermediate vector; adjusting and updating the end feature vector based on the node feature vector and a node mapping weight to obtain a second intermediate vector; wherein the node mapping weight refers to a mapping weight of a path between a path start node and a path end node completed by a target user number; splicing the first intermediate vector and the second intermediate vector to obtain the target feature vector.

3. The method of claim 1, wherein, The step of acquiring the resource operation feature vector corresponding to the resource operation feature set associated with the target user number comprises the following steps: acquiring account attribute information of a target user resource account registered by the target user number in the resource management platform, and historical resource operation records under each channel associated with the target user resource account; acquiring historical call records initiated by all user numbers before the target voice call; extract a channel feature corresponding to the target user resource account from the account attribute information and the historical resource operation record and the historical call record, to obtain the resource operation feature set; encode the resource operation feature set through a double-layer random forest model to obtain the resource operation feature vector.

4. The method of claim 1, wherein, The training of the initialized path prediction model based on the sample resource operation feature vector, the sample node feature vector and the sample terminal feature vector includes: performing compression transformation on the sample resource operation feature vector to obtain a first sample intermediate vector; adjusting and updating the sample terminal feature vector based on the sample node feature vector and the adjusted node mapping weight to obtain a second sample intermediate vector, wherein the node mapping weight is a mapping weight between the node of the non-path terminal and the path terminal; splicing the first sample intermediate vector and the second sample intermediate vector to obtain the sample feature vector; training the initialized path prediction model using the sample feature vector.

5. The method of claim 1, wherein, The training of the initialized path prediction model based on the sample resource operation feature vector, the sample node feature vector and the sample terminal feature vector includes: in the case where the training output result corresponding to the sample feature vector and the label matched with the sample feature vector do not match, the training error corresponding to the training output result is back propagated to adjust the training weight in the path prediction model.

6. A device for predicting a voice interaction path, characterized in that comprise: an acquisition unit, configured to acquire a resource operation feature vector corresponding to a resource operation feature set of a target user number in a case where a target voice call initiated by the target user number is received, wherein the resource operation feature set comprises resource operation features of the target user number acquired under each channel in a resource management platform; a determination unit, configured to determine a node feature vector corresponding to a pre-path node completed by the target user number in the target voice call, and a terminal feature vector corresponding to all path terminals associated with the target voice call, wherein the pre-path node is a path node corresponding to a voice interaction sub-service currently completed by the target user number; a prediction unit, configured to input a target feature vector to be predicted determined based on the resource operation feature vector, the node feature vector and the terminal feature vector into a path prediction model, wherein the path prediction model is used to calculate a prediction probability that each candidate path node is a target path node, wherein the target path node is a path node corresponding to a target voice interaction sub-service to be predicted next after the currently completed voice interaction sub-service for the target user number; and determine the target path node from the candidate path nodes according to a sorting result of the prediction probability. The device is also used for: obtaining sample data of sample voice calls initiated by a plurality of sample user numbers; determining a sample resource operation feature vector associated with the sample user numbers; encoding non-path end nodes in all path nodes of the sample voice calls to obtain sample node feature vectors, and encoding all path ends of the sample voice calls to obtain sample end feature vectors; and training an initialized path prediction model based on sample feature vectors determined by the sample resource operation feature vector, the sample node feature vectors, and the sample end feature vectors, until a convergence condition is reached.

7. A computer readable storage medium, characterized in that, The computer readable storage medium comprises a stored program, wherein the program is run by a processor to execute the method in any one of claims 1 to 5.

Images