Method and device for channel allocation according to user problems

Abstract

The embodiments of this specification provide a method and device for channel allocation for user problems. A specific implementation of the method includes: obtaining user questions asked by users; using a pre-trained manual satisfaction prediction model to predict the manual satisfaction of users with manual services after entering a first manual channel based on the user questions; using a pre-trained self-service resolution rate prediction model to predict the self-service resolution rate of user problems after the user enters a self-service channel based on the user questions; determining a target channel from multiple channels based on the manual satisfaction rate and the self-service resolution rate, and allocating the user problems to the target channel, wherein the multiple channels include at least a first manual channel and a self-service channel.

Description

Technical Field

[0001] The embodiments of this specification relate to the field of computer technology, and in particular, to a method and device for allocating channels according to user problems. Background Art

[0002] As the company's business continues to expand, how to provide customers with a high-quality customer service experience is a concern for most companies. However, on different dates (such as weekdays, weekends, holidays, etc.), or different time periods on the same day (such as daytime and night), the frequency of customers calling hotlines or using online services is different. Peak time periods will inevitably cause huge pressure on customer service personnel. If the scheduling is not good, it will greatly extend the user's waiting time, and even make the user's demands (also known as user problems) unable to be solved in a timely manner, which will greatly affect the user experience.

[0003] Therefore, there is an urgent need for a reasonable and reliable solution that can allocate user problems to appropriate channels, thereby improving the efficiency of customer service scheduling and enabling user problems to be solved in a timely manner. Summary of the invention

[0004] The embodiments of this specification provide a method and device for channel allocation for user problems, which can allocate user problems to appropriate channels, thereby improving the efficiency of customer service scheduling and enabling user problems to be solved in a timely manner.

[0005] In a first aspect, an embodiment of the present specification provides a method for channel allocation for user problems, including: obtaining user questions asked by users; using a pre-trained manual satisfaction prediction model to predict the manual satisfaction of the user with manual service after entering a first manual channel based on the user questions; using a pre-trained self-service resolution rate prediction model to predict the self-service resolution rate of the user problem after the user enters a self-service channel based on the user questions; determining a target channel from multiple channels based on the manual satisfaction rate and the self-service resolution rate, and allocating the user problem to the target channel, wherein the multiple channels include at least the first manual channel and the self-service channel.

[0006] In some embodiments, the target channel is determined from multiple channels based on the manual satisfaction rate and the self-service resolution rate, including: if the manual satisfaction rate reaches a satisfaction threshold and the self-service resolution rate does not reach a resolution rate threshold, the first manual channel is determined as the target channel; if the manual satisfaction rate does not reach the satisfaction threshold and the self-service resolution rate reaches the resolution rate threshold, the self-service channel is determined as the target channel; if the manual satisfaction rate reaches the satisfaction threshold and the self-service resolution rate reaches the resolution rate threshold, one of the first manual channel and the self-service channel is selected as the target channel.

[0007] In some embodiments, the selecting one of the first manual channel and the self-service channel as the target channel includes: obtaining the idleness information of the first manual channel; and selecting one of the first manual channel and the self-service channel as the target channel based on the idleness information.

[0008] In some embodiments, the selecting one of the first manual channel and the self-service channel as the target channel includes: determining whether the current moment is in a set busy time period; if the determination result is yes, selecting the self-service channel as the target channel.

[0009] In some embodiments, the multiple channels also include a second manual channel; and the target channel is determined from the multiple channels based on the manual satisfaction rate and the self-service resolution rate, including: if the manual satisfaction rate does not reach a satisfaction threshold and the self-service resolution rate does not reach a resolution rate threshold, then the second manual channel is determined as the target channel.

[0010] In some embodiments, the method further includes: obtaining behavior information and / or factor information related to the user problem; the behavior information includes first behavior data generated by the user in the target application to which the user problem belongs, and / or second behavior data generated by the target application for the user; the factor information includes a user factor of the user, a business factor related to the business to which the user problem belongs, and / or an experience factor of the user under the user problem; and using a pre-trained manual satisfaction prediction model to predict the manual satisfaction of the user with the manual service after entering the first manual channel according to the user problem, including: using the manual satisfaction prediction model to predict the manual satisfaction according to the user problem, and the behavior information and / or factor information; using a pre-trained self-service resolution rate prediction model to predict the self-service resolution rate of the user problem after the user enters the self-service channel according to the user problem, including: using the self-service resolution rate prediction model to predict the self-service resolution rate according to the user problem, and the behavior information and / or factor information.

[0011] In some embodiments, the artificial satisfaction prediction model includes multiple satisfaction prediction models; and using the artificial satisfaction prediction model to predict the artificial satisfaction according to the user questions, and the behavior information and / or factor information, includes: using the multiple satisfaction prediction models to predict multiple satisfaction sub-results according to the user questions, and the behavior information and / or factor information; determining the artificial satisfaction according to the multiple satisfaction sub-results.

[0012] In some embodiments, the multiple satisfaction prediction models include a first hybrid network, and at least one of the following: a first Bert model, a first XGBoost model.

[0013] In some embodiments, the use of the multiple satisfaction prediction models to predict multiple satisfaction sub-results according to the user questions, and the behavior information and / or factor information includes: using the first hybrid network to predict the first satisfaction sub-result according to the user questions, and the behavior information and / or factor information.

[0014] In some embodiments, the multiple satisfaction prediction models include the first Bert model; and the use of the multiple satisfaction prediction models to predict multiple satisfaction sub-results based on the user questions, and the behavior information and / or factor information also includes: using the first Bert model to predict a second satisfaction sub-result based on the user questions.

[0015] In some embodiments, the multiple satisfaction prediction models include the first XGBoost model; and the use of the multiple satisfaction prediction models to predict multiple satisfaction sub-results based on the user questions, and the behavior information and / or factor information also includes: using the first XGBoost model to predict a third satisfaction sub-result based on the factor information.

[0016] In some embodiments, the self-service resolution rate prediction model includes multiple resolution rate prediction models; and using the self-service resolution rate prediction model to predict the self-service resolution rate according to the user problem, and the behavior information and / or factor information, includes: using the multiple resolution rate prediction models to predict multiple resolution rate sub-results according to the user problem, and the behavior information and / or factor information; determining the self-service resolution rate according to the multiple resolution rate sub-results.

[0017] In some embodiments, the plurality of solution rate prediction models include a second hybrid network, and at least one of the following: a second Bert model, a second XGBoost model.

[0018] In some embodiments, the use of the multiple solution rate prediction models predicts multiple solution rate sub-results based on the user problem, and the behavior information and / or factor information, including: using the second hybrid network to predict a first solution rate sub-result based on the user problem, and the behavior information and / or factor information.

[0019] In some embodiments, the multiple solution rate prediction models include the second Bert model; and the use of the multiple solution rate prediction models to predict multiple solution rate sub-results based on the user questions, and the behavior information and / or factor information also includes: using the second Bert model to predict the second solution rate sub-result based on the user questions.

[0020] In some embodiments, the multiple solution rate prediction models include the second XGBoost model; and using the multiple solution rate prediction models to predict multiple solution rate sub-results based on the user questions, and the behavior information and / or factor information, also includes: using the second XGBoost model to predict a third solution rate sub-result based on the factor information.

[0021] In some embodiments, the first hybrid network is obtained by training an initial hybrid network using the following first training step: obtaining multiple first training samples, wherein any first training sample includes a first historical user question, an artificial satisfaction label associated with the first historical user question, and at least one of the following: behavior information, factor information; for a first training sample among the multiple first training samples, taking the first historical user question in the first training sample, and the behavior information and / or factor information associated with the first historical user question as input, and the artificial satisfaction label associated with the first historical user question as a training label, training the initial hybrid network to obtain a first hybrid network.

[0022] In some embodiments, the first Bert model is obtained by training the initial Bert model using the following second training step: for a first training sample among the multiple first training samples, the first historical user question in the first training sample is used as input, and the artificial satisfaction label associated with the first historical user question is used as a training label, and the initial Bert model is trained to obtain a first Bert model.

[0023] In some embodiments, the arbitrary first training sample includes factor information, and the first XGBoost model is obtained by training the initial XGBoost model using the following third training step: for a first training sample among the multiple first training samples, the factor information associated with the first historical user question in the first training sample is used as input, and the artificial satisfaction label associated with the first historical user question is used as a training label, and the initial XGBoost model is trained to obtain a first XGBoost model.

[0024] In some embodiments, the second hybrid network is obtained by training the initial hybrid network using the following fourth training step: obtaining multiple second training samples, wherein any second training sample includes a second historical user problem, and a self-service resolution rate label associated with the second historical user problem and at least one of the following: behavior information, factor information; for a second training sample among the multiple second training samples, the second historical user problem in the second training sample, and the behavior information and / or factor information associated with the second historical user problem are used as input, and the self-service resolution rate label associated with the second historical user problem is used as a training label to train the initial hybrid network to obtain a second hybrid network.

[0025] In some embodiments, the second Bert model is obtained by training the initial Bert model using the following fifth training step: for a second training sample among the multiple second training samples, the second historical user question in the second training sample is used as input, and the self-service resolution rate label associated with the second historical user question is used as a training label, and the initial Bert model is trained to obtain a second Bert model.

[0026] In some embodiments, the arbitrary second training sample includes factor information, and the second XGBoost model is obtained by training the initial XGBoost model using the following sixth training step: for a second training sample among the multiple second training samples, the factor information associated with the second historical user problem in the second training sample is used as input, and the self-service solution rate label associated with the second historical user problem is used as a training label to train the initial XGBoost model to obtain a second XGBoost model.

[0027] In a second aspect, an embodiment of the present specification provides a device for channel allocation for user problems, comprising: an acquisition unit, configured to acquire user questions asked by users; a manual satisfaction prediction unit, configured to use a pre-trained manual satisfaction prediction model to predict, based on the user questions, the manual satisfaction of the user with manual services after entering a first manual channel; a self-service resolution rate prediction unit, configured to use a pre-trained self-service resolution rate prediction model to predict, based on the user questions, the self-service resolution rate of the user problems after the user enters a self-service channel; a channel allocation unit, configured to determine a target channel from a plurality of channels based on the manual satisfaction rate and the self-service resolution rate, and allocate the user problems to the target channel, wherein the plurality of channels include at least the first manual channel and the self-service channel.

[0028] In a third aspect, an embodiment of the present specification provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed in a computer, the computer is caused to execute the method described in any implementation manner in the first aspect.

[0029] In a fourth aspect, an embodiment of the present specification provides a computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method described in any implementation manner in the first aspect is implemented.

[0030] In a fifth aspect, an embodiment of the present specification provides a computer program, wherein when the computer program is executed in a computer, the computer is caused to execute the method described in any implementation manner in the first aspect.

[0031] The above-mentioned embodiments of this specification provide a method and device for channel allocation for user problems. After obtaining the user questions asked by the user, the manual satisfaction of the user with the manual service after entering the first manual channel can be predicted based on the user questions, as well as the self-service resolution rate of the user problems after entering the self-service channel. Afterwards, a more suitable target channel can be determined from multiple channels with reference to the manual satisfaction rate and the self-service resolution rate, and the user problems can be allocated to the target channels, wherein the multiple channels at least include the first manual channel and the self-service channel. In this way, the user problems can be allocated to the appropriate channels, thereby improving the efficiency of customer service scheduling and enabling user problems to be resolved in a timely manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions of the multiple embodiments disclosed in this specification, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only the multiple embodiments disclosed in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0033] Figure 1 is an exemplary system architecture diagram to which some embodiments of the present specification may be applied;

[0034] Figure 2 It is a schematic diagram of the channel selection method;

[0035] Figure 3 is a flow chart of an embodiment of a method for channel allocation for user problems;

[0036] Figure 4 is a flow chart of another embodiment of a method for channel allocation for user problems;

[0037] Figure 5 It is a schematic diagram of the target channel determination process;

[0038] Figure 6 It is a structural diagram of a device for channel allocation according to user problems. DETAILED DESCRIPTION

[0039] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. The described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of this application.

[0040] It should be noted that, for the convenience of description, only the parts related to the invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments in this specification can be combined with each other.

[0041] As mentioned earlier, if customer service is not well dispatched, it will greatly extend the user's waiting time, and even make the user's demands unable to be resolved in a timely manner, which will greatly affect the user experience.

[0042] Based on this, some embodiments of this specification provide a method for channel allocation for user problems, through which user problems can be allocated to appropriate channels, thereby improving the efficiency of customer service scheduling and enabling user problems to be solved in a timely manner. Specifically, Figure 1 An exemplary system architecture diagram suitable for use with these embodiments is shown.

[0043] like Figure 1 As shown, the system architecture may include a server, which may belong to a customer service dispatch system, etc., and is not specifically limited here.

[0044] In practice, the server can obtain user questions asked by users, such as Figure 1 The user question A shown in FIG. The user question A may be a question asked by the user through online consultation or by calling a hotline, and is not specifically limited here. When the user question A is a question asked by calling a hotline, the user question A may be obtained by converting the user's voice information into text.

[0045] Afterwards, the server can predict the user's entry into the first manual channel (such as Figure 1 The figure shows the human satisfaction with manual service after entering the ordinary manual channel), and the self-service resolution rate of user problem A after the user enters the self-service channel.

[0046] Each manual customer service under the ordinary manual channel can usually handle most of the user's problems. Manual satisfaction can be specifically understood as the probability of whether the user is satisfied with the manual service after entering the ordinary manual channel. Under the self-service channel, users can use various self-service methods to solve problems. For example, they can use intelligent robots to answer questions with manual services, send SMS links, and / or go to the Q&A page of the APP (Application). The self-service resolution rate can be specifically understood as the probability of whether the user's problem can be solved after entering the self-service channel.

[0047] Next, the server can determine the target channel from multiple channels based on the predicted manual satisfaction and self-service resolution rate. The target channel is a channel that is more suitable for user problem A. For example, Figure 2 As shown, when manual satisfaction is high and the self-service resolution rate is low, the ordinary manual channel can be selected as the target channel. When manual satisfaction is low and the self-service resolution rate is high, the self-service channel can be selected as the target channel. When both manual satisfaction and self-service resolution rates are high, one channel can be selected from ordinary manual channels and self-service channels as the target channel. In practice, thresholds, such as satisfaction thresholds and resolution rate thresholds, can be used as a basis for measuring manual satisfaction and self-service resolution rates. It should be noted that the threshold can be defined manually or obtained by system statistics, and is not specifically limited here.

[0048] In some embodiments, when both the manual satisfaction rate and the self-service resolution rate are low, the user problem A can be given special attention. For example, the above-mentioned multiple channels can also include a second manual channel, such as Figure 2 The special manual channel shown in can be selected as the target channel when both the manual satisfaction and the self-service resolution rate are low.

[0049] The business capabilities of manual customer service in special manual channels are usually higher than those of manual customer service in ordinary manual channels. For user problems that cannot be solved through ordinary manual channels and self-service channels, by assigning the user problem to a special manual channel, manual customer service with stronger business capabilities can handle the user problem, which can improve the user problem resolution rate and user satisfaction.

[0050] After determining the target channel, the server can assign user problem A to the target channel. In this way, user problems can be assigned to appropriate channels, thereby improving the efficiency of customer service scheduling and enabling user problems to be solved in a timely manner.

[0051] The specific implementation steps of the above method are described below in conjunction with specific embodiments.

[0052] See also Figure 3 , which shows a process 300 of an embodiment of a method for channel allocation for user questions. The execution subject of the method can be Figure 1 The server side shown in the figure comprises the following steps:

[0053] Step 302, obtaining a user question asked by a user;

[0054] Step 304, using the pre-trained manual satisfaction prediction model, predicting the manual satisfaction of the user with the manual service after entering the first manual channel according to the user's question;

[0055] Step 306, using the pre-trained self-service resolution rate prediction model, predict the self-service resolution rate of the user's problem after the user enters the self-service channel according to the user's problem;

[0056] Step 308, determining a target channel from multiple channels based on manual satisfaction and self-service resolution rate, and assigning the user problem to the target channel, wherein the multiple channels at least include a first manual channel and a self-service channel.

[0057] The above steps are further explained below.

[0058] In step 302, the user questions asked by the user can be obtained in real time. For example, the user questions asked by the user through online consultation can be obtained. Alternatively, the user's voice information can be obtained and the voice information can be converted into text to obtain the user question. The voice information can be generated by calling a hotline.

[0059] Next, in step 304, the pre-trained manual satisfaction prediction model can be used to predict the manual satisfaction of the user to the manual service after entering the first manual channel according to the user's question. The first manual channel can include, for example, the ordinary manual channel as described above. The manual satisfaction can be a value within [0, 1].

[0060] As an example, the artificial satisfaction prediction model can be obtained by training the initial first machine learning model. The first machine learning model can be an untrained or untrained model. Further, the first machine learning model can include, for example, a Bert model, a hybrid network (MixNet) or a convolutional neural network (Convolutional Neural Networks, CNN), etc. Based on this, in step 304, the user question can be input into the artificial satisfaction prediction model, so that the artificial satisfaction prediction model predicts and outputs the artificial satisfaction.

[0061] It should be noted that when training the first machine learning model, the first machine learning model can be trained using the pre-collected multiple first historical user questions as inputs, and the artificial satisfaction labels associated with the multiple first historical user questions as training labels, thereby obtaining an artificial satisfaction prediction model. Specifically, during the training process, the prediction loss can be determined based on the artificial satisfaction labels as training labels and the prediction results of the first machine learning model, and the network parameters in the first machine learning model can be adjusted with the goal of reducing the prediction loss.

[0062] The first historical user problem may be a user problem solved through the first manual channel. In practice, after the user ends the conversation with the manual customer service under the first manual channel, the user may evaluate the satisfaction level of the manual customer service. The evaluated satisfaction level may be included in multiple satisfaction levels, and the multiple satisfaction levels may be, from high to low, very satisfied, satisfied, average, dissatisfied, and very dissatisfied. Based on this, the manual satisfaction label associated with the first historical user problem may be determined according to the satisfaction level associated with the first historical user problem.

[0063] As an example, when the satisfaction level of the first historical user question is very satisfied or satisfied, for example, the value 1 can be set as the artificial satisfaction label of the first historical user question; when the satisfaction level of the first historical user question is average, dissatisfied or very dissatisfied, for example, the value 0 can be set as the artificial satisfaction label of the first historical user question.

[0064] As another example, when the satisfaction level of the first historical user question is very satisfied, for example, a value 1 can be set as the artificial satisfaction label of the first historical user question; when the satisfaction level of the first historical user question is very unsatisfactory, for example, a value 0 can be set as the artificial satisfaction label of the first historical user question; when the satisfaction level of the first historical user question is satisfied, average or unsatisfactory, for example, a value 0.5 can be set as the artificial satisfaction label of the first historical user question.

[0065] It should be understood that the manual satisfaction label can be set according to actual conditions and is not specifically limited here.

[0066] Next, in step 306, the pre-trained self-service resolution rate prediction model can be used to predict the self-service resolution rate of the user's problem after the user enters the self-service channel according to the user's problem. The self-service resolution rate can be a value within [0, 1].

[0067] As an example, the self-service resolution rate prediction model can be obtained by training an initial second machine learning model. The second machine learning model can be an untrained or incompletely trained model. Furthermore, the second machine learning model can include, for example, a Bert model, a hybrid network, or a convolutional neural network. Based on this, in step 306, the user question can be input into the self-service resolution rate prediction model, so that the self-service resolution rate prediction model predicts and outputs the self-service resolution rate.

[0068] It should be noted that when training the second machine learning model, the plurality of pre-collected second historical user questions can be used as inputs, and the self-help resolution rate labels associated with the plurality of second historical user questions can be used as training labels to train the second machine learning model, thereby obtaining a self-help resolution rate prediction model. Specifically, during the training process, the prediction loss can be determined based on the self-help resolution rate label as the training label and the prediction result of the second machine learning model, and the network parameters in the second machine learning model can be adjusted with the goal of reducing the prediction loss.

[0069] The second historical user problem may be a user problem that is solved through a self-service channel. In practice, the user problem that is solved through a self-service channel is usually solved or unsolved. Based on this, the self-service resolution rate label associated with the second historical user problem can be determined according to the resolution status of the second historical user problem. For example, when the resolution status of the second historical user problem is solved, for example, a value of 1 can be set as the self-service resolution rate label of the second historical user problem, and when the resolution status of the second historical user problem is unsolved, for example, a value of 0 can be set as the self-service resolution rate label of the second historical user problem.

[0070] Next, in step 308, a target channel may be determined from multiple channels based on manual satisfaction and self-service resolution rate, and the user problem may be allocated to the target channel, wherein the multiple channels include at least a first manual channel and a self-service channel.

[0071] Specifically, if the manual satisfaction reaches the satisfaction threshold and the self-service resolution rate does not reach the resolution rate threshold, the first manual channel can be determined as the target channel. If the manual satisfaction does not reach the satisfaction threshold and the self-service resolution rate reaches the resolution rate threshold, the self-service channel can be determined as the target channel. If the manual satisfaction reaches the satisfaction threshold and the self-service resolution rate reaches the resolution rate threshold, one of the first manual channel and the self-service channel can be selected as the target channel.

[0072] Furthermore, when selecting one channel from the first manual channel and the self-service channel as the target channel, various selection methods may be used.

[0073] In one example, one channel may be randomly selected from the first manual channel and the self-service channel as the target channel.

[0074] In another example, the idleness information of the first manual channel can be obtained, and according to the idleness information, one of the first manual channel and the self-service channel can be selected as the target channel. For example, according to the idleness information, it can be determined whether there is an idle manual customer service in the first manual channel. If the determination result is yes, the first manual channel is determined as the target channel. If the determination result is no, the self-service channel is determined as the target channel. In this way, when there is an idle manual customer service in the first manual channel, the user problem can be preferentially assigned to the first manual channel to ensure that the user problem is solved in a timely manner.

[0075] In another example, it can be determined whether the current time is in a set busy time period. If the result is yes, the self-service channel can be selected as the target channel. In practice, if a user seeks help from a human customer service representative during a busy time period, they usually need to wait in line, and the user's request may not be resolved in a timely manner. By adopting this implementation method, user problems can be preferentially assigned to self-service channels during busy time periods, ensuring that user problems are resolved in a timely manner.

[0076] In some embodiments, the above-mentioned multiple channels may further include a second manual channel, which may include, for example, the special manual channel described above. In step 308, if the manual satisfaction rate does not reach the satisfaction threshold and the self-service resolution rate does not reach the resolution rate threshold, the second manual channel may be determined as the target channel.

[0077] In practice, the business capabilities of manual customer service in the second manual channel are usually higher than those in the first manual channel. For user problems that cannot be solved through the first manual channel and the self-service channel, by assigning the user problem to the second manual channel, manual customer service with stronger business capabilities can handle the user problem, which can improve the problem solving rate and user satisfaction.

[0078] In some embodiments, in step 308, if the manual satisfaction rate does not reach the satisfaction threshold and the self-service resolution rate does not reach the resolution rate threshold, the user problem can be assigned to the first manual channel and the self-service channel. In this way, the user problem can be handled in a combination of the first manual channel and the self-service channel, which can improve the resolution rate of the user problem and the user's satisfaction.

[0079] Figure 3In a corresponding embodiment, after obtaining the user question asked by the user, the manual satisfaction of the user with the manual service after entering the first manual channel and the self-service resolution rate of the user question after entering the self-service channel can be predicted based on the user question. Afterwards, a more suitable target channel can be determined from multiple channels with reference to the manual satisfaction rate and the self-service resolution rate, and the user question can be assigned to the target channel, wherein the multiple channels at least include the first manual channel and the self-service channel. In this way, the user question can be assigned to the appropriate channel, thereby improving the efficiency of customer service scheduling and enabling the user problem to be resolved in a timely manner.

[0080] See further Figure 4 , which shows a process 400 of another embodiment of a method for channel allocation for user questions. The execution subject of the method may be as follows Figure 1 The server side shown in the figure comprises the following steps:

[0081] Step 402, obtaining a user question asked by the user, and behavior information and / or factor information related to the user question;

[0082] Step 404, using the pre-trained manual satisfaction prediction model, predicting the manual satisfaction of the user with the manual service after entering the first manual channel according to the user's question and the behavior information and / or factor information related to the user's question;

[0083] Step 406, using the pre-trained self-service resolution rate prediction model, based on the user's problem and the behavior information and / or factor information related to the user's problem, predict the self-service resolution rate of the user's problem after the user enters the self-service channel;

[0084] Step 408, determining a target channel from multiple channels based on manual satisfaction and self-service resolution rate, and assigning the user problem to the target channel, wherein the multiple channels at least include a first manual channel and a self-service channel.

[0085] The above steps are further explained below.

[0086] In step 402, the user question asked by the user and the behavior information and / or factor information related to the user question can be obtained in real time. Here, for the explanation of the user question, please refer to the relevant description in the previous text, which will not be repeated here.

[0087] The behavior information related to the user question may include, for example, the first behavior data generated by the user in the target application to which the user question belongs, and / or the second behavior data generated by the target application for the user. The target application may specifically be the application where the online consultation portal or hotline portal used by the user to ask the user question is located, or the application associated with the user identifier provided by the user when asking the user question, etc., which are not specifically limited here. In addition, the target application can be any category of application, such as payment applications, social applications, game applications, or educational applications, etc.

[0088] The first behavior data may specifically include the behavior data generated by the user in the target application in the recent period, such as in the recent few days (such as the recent 2, 3, 4 or 5 days, etc.) or in the recent preset number of times (such as the recent 30, 40 or 50 times, etc.), and the behavior data may include, for example, click behavior data and / or browsing behavior data, etc. The second behavior data may specifically include the behavior data generated by the target application for the user in the recent period, and the behavior data may include, for example, information actively pushed to the user by the target application, such as bills, reminder information and / or advertisements actively pushed to the user. It should be noted that the first behavior data and the second behavior data may be respectively embodied as behavior sequences, and each piece of data in the behavior sequence may be sorted in chronological order.

[0089] The factor information related to the user question may include, for example, a user factor of the user, a business factor related to the business to which the user question belongs, and / or an experience factor of the user under the user question.

[0090] User factors may include, for example, the user's age, gender, and / or familiarity with the target application, etc. The familiarity with the target application may be represented, for example, by the total number of times the user has logged into the target application, the number of times the user has recently logged into the target application, the number of years the user has used the target application, or the time when the user last logged into the target application, etc.

[0091] The business factor may include, for example, the business identifier of the business to which the user question belongs, the scenario identifier of the business scenario to which the business belongs, and the historical satisfaction and / or historical resolution rate corresponding to the target question. The target question is the question corresponding to the user question, and the question may be included in the question library. It should be noted that the question may be an abbreviation of a standard question, and a plurality of standard questions may be stored in the question library.

[0092] In practice, the historical satisfaction level corresponding to the target question can be determined based on the manual satisfaction level of multiple first historical user questions corresponding to the target question in the first manual channel. The historical resolution rate corresponding to the target question can be determined based on the self-service resolution rate of multiple second historical user questions corresponding to the target question in the self-service channel. It should be understood that the historical satisfaction level can be used as a reference factor when predicting the manual satisfaction level, and the historical resolution rate can be used as a reference factor when predicting the self-service resolution rate.

[0093] Experience factors may include, for example, the number of interaction rounds, intent recognition results, the number of repeated requests for help, and / or the number of repeated visits, etc.

[0094] Among them, the number of interaction rounds can be understood as the number of rounds of interaction between the user and the first intelligent robot for the user problem before the channel is assigned to the user problem. It should be pointed out that the first intelligent robot can be a robot used for preliminary processing of user problems, and the first intelligent robot is usually different from the intelligent robot under the self-service channel. In practice, before the user problem is assigned to a channel, the user may have several rounds of interaction with the first intelligent robot. For example, a hotline service and / or an online consultation service may have its own first intelligent robot. When a user calls a hotline or conducts an online consultation, the corresponding first intelligent robot can, for example, ask the user for information so that the user can answer. Among them, the information that the first intelligent robot asks the user may include, but is not limited to, the type of business to which the user problem that the user needs to ask belongs, and / or the content of the user problem, etc., which are not specifically limited here.

[0095] The intention recognition result can be understood as the result of the first intelligent robot recognizing the user's intention according to the user's question. The number of repeated requests for help can be understood as the number of times the user repeatedly calls the hotline for the user's question. The number of repeated visits can be understood as the number of times the user repeatedly visits the target application for the user's question. In other words, the number of repeated visits can be the number of times the user repeatedly consults online for the user's question.

[0096] It should be noted that the behavior information and factor information described above can be used as auxiliary data for predicting manual satisfaction and self-service resolution rate. This auxiliary data helps to improve the accuracy of the final predicted manual satisfaction and self-service resolution rate.

[0097] Next, in step 404, the pre-trained manual satisfaction prediction model can be used to predict the manual satisfaction of the user with the manual service after entering the first manual channel according to the user's question and the behavior information and / or factor information related to the user's question. The manual satisfaction can be a value within [0, 1].

[0098] As an example, the artificial satisfaction prediction model can be obtained by training an initial hybrid network. The initial hybrid network can be an untrained or untrained model. Based on this, user problems, as well as behavioral information and / or factor information related to user problems can be input into the artificial satisfaction prediction model, so that the artificial satisfaction prediction model predicts and outputs artificial satisfaction based on the input information. When inputting factor information into the artificial satisfaction prediction model, the factor information can be normalized first, and then the normalized factor information can be input into the artificial satisfaction prediction model.

[0099] It should be noted that when training the initial hybrid network, multiple first training samples can be obtained, wherein any first training sample can include a first historical user question, an artificial satisfaction label associated with the first historical user question, and at least one of the following: behavior information, factor information. For the first training sample among the multiple first training samples, the first historical user question in the first training sample, and the behavior information and / or factor information associated with the first historical user question can be used as input, and the artificial satisfaction label associated with the first historical user question can be used as a training label to train the initial hybrid network to obtain an artificial satisfaction prediction model. Specifically, during the training process, the prediction loss can be determined based on the artificial satisfaction label as the training label and the prediction result of the initial hybrid network, and the network parameters in the initial hybrid network can be adjusted with the goal of reducing the prediction loss.

[0100] Next, in step 406, the self-service resolution rate prediction model trained in advance can be used to predict the self-service resolution rate of the user's problem after the user enters the self-service channel according to the user's problem and the behavior information and / or factor information related to the user's problem. The self-service resolution rate can be a value within [0, 1].

[0101] As an example, the self-service resolution rate prediction model can be obtained by training an initial hybrid network. The initial hybrid network can be an untrained or incompletely trained model. Based on this, the user problem, as well as the behavior information and / or factor information related to the user problem, can be input into the self-service resolution rate prediction model, so that the self-service resolution rate prediction model predicts and outputs the self-service resolution rate.

[0102] It should be noted that when training the initial hybrid network, multiple second training samples can be obtained, wherein any second training sample can include a second historical user problem, a self-service resolution rate label associated with the second historical user problem, and at least one of the following: behavior information, factor information. For a second training sample among the multiple second training samples, the second historical user problem in the second training sample, and the behavior information and / or factor information associated with the second historical user problem can be used as input, and the self-service resolution rate label associated with the second historical user problem can be used as a training label to train the initial hybrid network to obtain a self-service resolution rate prediction model. Specifically, during the training process, the prediction loss can be determined based on the self-service resolution rate label as the training label and the prediction result of the initial hybrid network, and the network parameters in the initial hybrid network can be adjusted with the goal of reducing the prediction loss.

[0103] It should be noted that when predicting manual satisfaction, the business factors used may include the historical satisfaction corresponding to the target question. When predicting the self-service resolution rate, the business factors used may include the historical resolution rate corresponding to the target question. For example, in step 404, if the factor information is used as input, and the factor information includes a business factor, the business factor may include, for example, the historical satisfaction corresponding to the target question. In step 406, if the factor information is used as input, and the factor information includes a business factor, the business factor may include, for example, the historical resolution rate corresponding to the target question.

[0104] Next, in step 408, a target channel may be determined from multiple channels based on manual satisfaction and self-service resolution rate, and the user problem may be assigned to the target channel, wherein the multiple channels include at least the first manual channel and the self-service channel. Figure 3 For the similarity of step 308 in the corresponding embodiment, reference may be made to the relevant description of step 308, which will not be repeated here.

[0105] Figure 4 In the corresponding embodiment, the accuracy of the final prediction result can be further improved by using the user questions asked by the user, as well as the behavior information and / or factor information related to the user questions as reference data for predicting the manual satisfaction and self-service resolution rate. In this way, the user questions can be assigned to more appropriate channels, thereby further improving the efficiency of customer service scheduling and enabling user problems to be resolved in a timely manner.

[0106] exist Figure 4In the corresponding embodiment, as an implementation method, the manual satisfaction prediction model may include multiple satisfaction prediction models, and step 404 may further include: using the multiple satisfaction prediction models to predict multiple satisfaction sub-results based on user questions, and behavior information and / or factor information related to the user questions; and determining the user's manual satisfaction with the manual service after entering the first manual channel based on the multiple satisfaction sub-results.

[0107] In practice, the above-mentioned multiple satisfaction prediction models may include, for example, a first hybrid network and at least one of the following: a first Bert model, a first XGBoost model, etc. Figure 5 As shown, it shows that the above-mentioned multiple satisfaction prediction models include a first hybrid network, a first Bert model and a first XGBoost model. Among them, Figure 5 It is a schematic diagram of the target channel determination process.

[0108] like Figure 5 As shown, the first hybrid network can be used to predict the first satisfaction sub-result based on the user question and the behavior information and / or factor information related to the user question. For example, the user question and the behavior information and / or factor information related to the user question can be input into the first hybrid network, so that the first hybrid network predicts and outputs the first satisfaction sub-result.

[0109] In addition, when the above-mentioned multiple satisfaction prediction models include the first Bert model, the first Bert model can also be used to predict the second satisfaction sub-result according to the user question. For example, the user question can be input into the first Bert model, so that the first Bert model predicts and outputs the second satisfaction sub-result.

[0110] In addition, in the case of obtaining factor information related to the user question, when the above-mentioned multiple satisfaction prediction models include the first XGBoost model, the first XGBoost model can also be used to predict the third satisfaction sub-result according to the factor information. For example, the factor information can be input into the first XGBoost model, so that the first XGBoost model predicts and outputs the third satisfaction sub-result.

[0111] It should be understood that the prediction results of the first hybrid network, the first Bert model, and the first XGBoost model can all be referred to as artificial satisfaction. In order to distinguish it from the final predicted artificial satisfaction, in the embodiment of this specification, the prediction result is referred to as a satisfaction sub-result.

[0112] In practice, the first hybrid network, the first Bert model and the first XGBoost model can be trained separately.

[0113] For example, the first hybrid network can be obtained by training the initial hybrid network using the following first training step: obtaining multiple first training samples, wherein any first training sample includes a first historical user question, an artificial satisfaction label associated with the first historical user question, and at least one of the following: behavior information, factor information; for the first training sample among the multiple first training samples, the first historical user question in the first training sample, and the behavior information and / or factor information associated with the first historical user question are used as inputs, and the artificial satisfaction label associated with the first historical user question is used as a training label, and the initial hybrid network is trained to obtain the first hybrid network. The initial hybrid network may be an untrained or uncompleted model. Specifically, during the training process, the prediction loss may be determined based on the artificial satisfaction label as the training label and the prediction result of the initial hybrid network, and the network parameters in the initial hybrid network may be adjusted with the goal of reducing the prediction loss.

[0114] The first Bert model can be obtained by training the initial Bert model using the following second training step: for the first training sample among the above-mentioned multiple first training samples, the first historical user question in the first training sample is used as input, and the artificial satisfaction label associated with the first historical user question is used as a training label, and the initial Bert model is trained to obtain the first Bert model. The initial Bert model may be an untrained or uncompleted model. Specifically, during the training process, the prediction loss can be determined based on the artificial satisfaction label as the training label and the prediction result of the initial Bert model, and the network parameters in the initial Bert model are adjusted with the goal of reducing the prediction loss.

[0115] When any first training sample includes factor information, the first XGBoost model can be obtained by training the initial XGBoost model using the following third training step: for the first training sample among the above-mentioned multiple first training samples, the factor information associated with the first historical user question in the first training sample is used as input, and the artificial satisfaction label associated with the first historical user question is used as a training label, and the initial XGBoost model is trained to obtain the first XGBoost model. The initial XGBoost model may be an untrained or uncompleted model. Specifically, during the training process, the prediction loss can be determined based on the artificial satisfaction label as the training label and the prediction result of the initial XGBoost model, and the network parameters in the initial XGBoost model are adjusted with the goal of reducing the prediction loss.

[0116] After predicting multiple satisfaction sub-results, we can Figure 5As shown in , according to the multiple satisfaction sub-results, the manual satisfaction of the user with the manual service after entering the first manual channel is determined. For example, the average value of the multiple satisfaction sub-results can be determined as the manual satisfaction. For another example, when the multiple satisfaction prediction models are respectively provided with weight values, the multiple satisfaction sub-results can be weighted and summed according to the weight values, and the obtained sum value can be determined as the manual satisfaction.

[0117] exist Figure 4 In a corresponding embodiment, as an implementation method, the self-service resolution rate prediction model may include multiple resolution rate prediction models, and step 406 may further include: using the multiple resolution rate prediction models to predict multiple resolution rate sub-results based on user problems, and behavioral information and / or factor information related to the user problems; and determining the self-service resolution rate of user problems after the user enters the self-service channel based on the multiple resolution rate sub-results.

[0118] In practice, the above-mentioned multiple solution rate prediction models may include, for example, a second hybrid network and at least one of the following: a second Bert model, a second XGBoost model, etc. Figure 5 As shown, it shows that the above-mentioned multiple solution rate prediction models include a second hybrid network, a second Bert model and a second XGBoost model.

[0119] like Figure 5 As shown, the second hybrid network can be used to predict the first resolution rate sub-result based on the user problem and the behavior information and / or factor information related to the user problem. For example, the user problem and the behavior information and / or factor information related to the user problem can be input into the second hybrid network, so that the second hybrid network predicts and outputs the first resolution rate sub-result.

[0120] In addition, when the above-mentioned multiple resolution rate prediction models include a second Bert model, the second Bert model can also be used to predict the first resolution rate according to the user question. For example, the user question can be input into the second Bert model, so that the second Bert model predicts and outputs the second resolution rate sub-result.

[0121] In addition, when the factor information related to the user's question is obtained, when the above-mentioned multiple solution rate prediction models include a second XGBoost model, the second XGBoost model can also be used to predict the first solution rate according to the factor information. For example, the factor information can be input into the second XGBoost model, so that the second XGBoost model predicts and outputs the third solution rate sub-result.

[0122] It should be understood that the prediction results of the second hybrid network, the second Bert model, and the second XGBoost model can all be referred to as self-help solution rates. In order to distinguish from the final predicted self-help solution rate, in the embodiment of this specification, the prediction result is referred to as the solution rate sub-result.

[0123] In practice, the second hybrid network, the second Bert model, and the second XGBoost model can be trained separately.

[0124] For example, the second hybrid network can be obtained by training the initial hybrid network using the following fourth training step: obtaining multiple second training samples, wherein any second training sample includes a second historical user problem, and a self-service resolution rate label associated with the second historical user problem and at least one of the following: behavior information, factor information; for a second training sample among the multiple second training samples, the second historical user problem in the second training sample, and the behavior information and / or factor information associated with the second historical user problem are used as input, and the self-service resolution rate label associated with the second historical user problem is used as a training label, and the initial hybrid network is trained to obtain a second hybrid network. The initial hybrid network may be an untrained or uncompleted model. Specifically, during the training process, the prediction loss may be determined based on the self-service resolution rate label as the training label and the prediction result of the initial hybrid network, and the network parameters in the initial hybrid network may be adjusted with the goal of reducing the prediction loss.

[0125] The second Bert model can be obtained by training the initial Bert model using the following fifth training step: for the second training sample in the above-mentioned multiple second training samples, the second historical user question in the second training sample is used as input, and the self-service resolution rate label associated with the second historical user question is used as a training label, and the initial Bert model is trained to obtain the second Bert model. Among them, the initial Bert model can be an untrained or uncompleted model. Specifically, during the training process, the prediction loss can be determined based on the self-service resolution rate label as the training label and the prediction result of the initial hybrid network, and the network parameters in the initial Bert model are adjusted with the goal of reducing the prediction loss.

[0126] When any second training sample includes factor information, the second XGBoost model can be obtained by training the initial XGBoost model using the following sixth training step: for a second training sample among the above-mentioned multiple second training samples, the factor information associated with the second historical user question in the second training sample is used as input, and the self-help solution rate label associated with the second historical user question is used as a training label, and the initial XGBoost model is trained to obtain a second XGBoost model. The initial XGBoost model may be an untrained or uncompleted model. Specifically, during the training process, the prediction loss can be determined based on the self-help solution rate label as the training label and the prediction result of the initial XGBoost model, and the network parameters in the initial XGBoost model are adjusted with the goal of reducing the prediction loss.

[0127] After predicting multiple solution rate sub-results, you can Figure 5 As shown in , the self-service resolution rate of the user problem after the user enters the self-service channel is determined based on the multiple resolution rate sub-results. For example, the average value of the multiple resolution rate sub-results can be determined as the self-service resolution rate. For another example, when the multiple resolution rate prediction models are respectively provided with weight values, the multiple resolution rate sub-results can be weighted and summed according to the weight values, and the obtained sum value can be determined as the self-service resolution rate.

[0128] Continue to view Figure 5 After obtaining the final manual satisfaction using the above multiple satisfaction prediction models and obtaining the final self-service resolution rate using the above multiple resolution rate prediction models, the target channel can be determined from the multiple channels mentioned above according to the manual satisfaction and the self-service resolution rate. Figure 3 The relevant description of step 308 in the corresponding embodiment will not be repeated here.

[0129] It should be noted that for the three models, the Bert model, the hybrid network and the XGBoost model, since the Bert model has a better prediction effect on text, the XGBoost model has a better prediction effect on factors, and the hybrid network can make predictions based on a variety of information, the above-mentioned multiple satisfaction prediction models and the above-mentioned multiple solution rate prediction models are all obtained by training the initial three models, and the prediction results of the three models can be combined to obtain a more accurate final prediction result.

[0130] Further references Figure 6 This specification provides an embodiment of a device for channel allocation according to user problems. The device can be applied to Figure 1 The server side is shown.

[0131] like Figure 6 As shown, the device 600 for channel allocation for user problems in this embodiment includes: an acquisition unit 601, a manual satisfaction prediction unit 602, a self-service resolution rate prediction unit 603 and a channel allocation unit 604. The acquisition unit 601 is configured to acquire user questions asked by users; the manual satisfaction prediction unit 602 is configured to use a pre-trained manual satisfaction prediction model to predict the manual satisfaction of users with manual services after entering the first manual channel according to user questions; the self-service resolution rate prediction unit 603 is configured to use a pre-trained self-service resolution rate prediction model to predict the self-service resolution rate of user problems after entering the self-service channel according to user questions; the channel allocation unit 604 is configured to determine a target channel from multiple channels according to the manual satisfaction rate and the self-service resolution rate, and allocate the user problem to the target channel, wherein the multiple channels at least include the first manual channel and the self-service channel.

[0132] In some embodiments, the channel allocation unit 604 can be further configured to: if the manual satisfaction rate reaches the satisfaction threshold and the self-service resolution rate does not reach the resolution rate threshold, the first manual channel is determined as the target channel; if the manual satisfaction rate does not reach the satisfaction threshold and the self-service resolution rate reaches the resolution rate threshold, the self-service channel is determined as the target channel; if the manual satisfaction rate reaches the satisfaction threshold and the self-service resolution rate reaches the resolution rate threshold, one of the first manual channel and the self-service channel is selected as the target channel.

[0133] In some embodiments, the channel allocation unit 604 may be further configured to: obtain the idleness information of the first manual channel; and select one of the first manual channel and the self-service channel as the target channel according to the idleness information.

[0134] In some embodiments, the channel allocation unit 604 may be further configured to: determine whether the current time is in a set busy time period; if the determination result is yes, select the self-service channel as the target channel.

[0135] In some embodiments, the above-mentioned multiple channels may also include a second manual channel; and the channel allocation unit 604 may be further configured to: if the manual satisfaction rate does not reach the satisfaction threshold and the self-service resolution rate does not reach the resolution rate threshold, the second manual channel will be determined as the target channel.

[0136] In some embodiments, the acquisition unit 601 can also be configured to: acquire behavior information and / or factor information related to user problems; the behavior information may include first behavior data generated by the user in the target application to which the user problem belongs, and / or second behavior data generated by the target application for the user; the factor information may include the user factor of the user, the business factor related to the business to which the user problem belongs, and / or the experience factor of the user under the user problem; and the manual satisfaction prediction unit 602 can be further configured to: use the manual satisfaction prediction model to predict manual satisfaction based on user problems and the above-mentioned behavior information and / or factor information; the self-service solution rate prediction unit 603 can be further configured to: use the self-service solution rate prediction model to predict the self-service solution rate based on user problems and the above-mentioned behavior information and / or factor information.

[0137] In some embodiments, the artificial satisfaction prediction model may include multiple satisfaction prediction models; and the artificial satisfaction prediction unit 602 may be further configured to: use the multiple satisfaction prediction models to predict multiple satisfaction sub-results based on user questions, and the above-mentioned behavior information and / or factor information; and determine the artificial satisfaction based on the multiple satisfaction sub-results.

[0138] In some embodiments, the above-mentioned multiple satisfaction prediction models may include a first hybrid network, and at least one of the following: a first Bert model, a first XGBoost model.

[0139] In some embodiments, the manual satisfaction prediction unit 602 may be further configured to: utilize the first hybrid network to predict a first satisfaction sub-result based on the user question and the above-mentioned behavior information and / or factor information.

[0140] In some embodiments, the above-mentioned multiple satisfaction prediction models may include a first Bert model; and the manual satisfaction prediction unit 602 may be further configured to: use the first Bert model to predict a second satisfaction sub-result based on user questions.

[0141] In some embodiments, the above-mentioned multiple satisfaction prediction models may include a first XGBoost model; and the manual satisfaction prediction unit 602 may be further configured to: use the first XGBoost model to predict a third satisfaction sub-result based on factor information.

[0142] In some embodiments, the self-service resolution rate prediction model may include multiple resolution rate prediction models; and the self-service resolution rate prediction unit 603 may be further configured to: utilize the multiple resolution rate prediction models to predict multiple resolution rate sub-results based on user questions, and the above-mentioned behavior information and / or factor information; and determine the self-service resolution rate based on the multiple resolution rate sub-results.

[0143] In some embodiments, the above-mentioned multiple solution rate prediction models may include a second hybrid network, and at least one of the following: a second Bert model, a second XGBoost model.

[0144] In some embodiments, the self-service resolution rate prediction unit 603 may be further configured to: utilize the second hybrid network to predict the first resolution rate sub-result according to the user question and the above-mentioned behavior information and / or factor information.

[0145] In some embodiments, the above-mentioned multiple solution rate prediction models may include a second Bert model; and the self-service solution rate prediction unit 603 may be further configured to: use the second Bert model to predict a second solution rate sub-result based on the user question.

[0146] In some embodiments, the above-mentioned multiple solution rate prediction models may include a second XGBoost model; and the self-service solution rate prediction unit 603 may be further configured to: use the second XGBoost model to predict the third solution rate sub-result based on the factor information.

[0147] In some embodiments, the first hybrid network can be obtained by training the initial hybrid network using the following first training step: obtaining multiple first training samples, wherein any first training sample includes a first historical user question, an artificial satisfaction label associated with the first historical user question, and at least one of the following: behavior information, factor information; for a first training sample among the multiple first training samples, taking the first historical user question in the first training sample, and the behavior information and / or factor information associated with the first historical user question as input, and the artificial satisfaction label associated with the first historical user question as a training label, training the initial hybrid network, and obtaining the first hybrid network.

[0148] In some embodiments, the first Bert model is obtained by training the initial Bert model using the following second training step: for a first training sample among the above-mentioned multiple first training samples, the first historical user question in the first training sample is used as input, and the artificial satisfaction label associated with the first historical user question is used as a training label, and the initial Bert model is trained to obtain the first Bert model.

[0149] In some embodiments, any first training sample includes factor information, and the first XGBoost model can be obtained by training the initial XGBoost model using the following third training step: for a first training sample among the above-mentioned multiple first training samples, the factor information associated with the first historical user question in the first training sample is used as input, and the artificial satisfaction label associated with the first historical user question is used as a training label, and the initial XGBoost model is trained to obtain a first XGBoost model.

[0150] In some embodiments, the second hybrid network can be obtained by training the initial hybrid network using the following fourth training step: obtaining multiple second training samples, wherein any second training sample includes a second historical user problem, a self-service resolution rate label associated with the second historical user problem, and at least one of the following: behavior information, factor information; for a second training sample among the multiple second training samples, taking the second historical user problem in the second training sample, and the behavior information and / or factor information associated with the second historical user problem as input, and the self-service resolution rate label associated with the second historical user problem as a training label, training the initial hybrid network, and obtaining a second hybrid network.

[0151] In some embodiments, the second Bert model can be obtained by training the initial Bert model using the following fifth training step: for a second training sample among the above-mentioned multiple second training samples, the second historical user question in the second training sample is used as input, and the self-service solution rate label associated with the second historical user question is used as a training label, and the initial Bert model is trained to obtain a second Bert model.

[0152] In some embodiments, any second training sample includes factor information, and the second XGBoost model can be obtained by training the initial XGBoost model using the following sixth training step: for a second training sample among the above-mentioned multiple second training samples, the factor information associated with the second historical user problem in the second training sample is used as input, and the self-service solution rate label associated with the second historical user problem is used as a training label, and the initial XGBoost model is trained to obtain a second XGBoost model.

[0153] exist Figure 6 In the corresponding device embodiment, the specific processing of each unit and the technical effects it brings can be referred to the relevant description of the method embodiment in the previous text, and will not be repeated here.

[0154] The embodiments of this specification also provide a computer-readable storage medium on which a computer program is stored. When the computer program is executed in a computer, the computer is caused to execute the method for allocating channels for user problems described in the above method embodiments.

[0155] The embodiments of the present specification also provide a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method of channel allocation for user problems described in the above method embodiments is implemented.

[0156] The embodiments of this specification also provide a computer program, wherein when the computer program is executed in a computer, the computer is caused to execute the method for allocating channels for user problems described in the above method embodiments.

[0157] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the multiple embodiments disclosed in this specification can be implemented by hardware, software, firmware, or any combination thereof. When implemented by software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0158] In some cases, the actions or steps described in the claims may be performed in a different order than in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0159] The specific implementation methods described above further illustrate in detail the purposes, technical solutions and beneficial effects of the multiple embodiments disclosed in this specification. It should be understood that the above description is only the specific implementation methods of the multiple embodiments disclosed in this specification, and is not used to limit the protection scope of the multiple embodiments disclosed in this specification. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solutions of the multiple embodiments disclosed in this specification should be included in the protection scope of the multiple embodiments disclosed in this specification.

Claims

1. A method for channel allocation for user problems, comprising: Get the user questions asked by the user; Using a pre-trained manual satisfaction prediction model, predicting the manual satisfaction of the user with the manual service after entering the first manual channel according to the user's question; Using a pre-trained self-service resolution rate prediction model, predict the self-service resolution rate of the user problem after the user enters the self-service channel according to the user problem; If the manual satisfaction rate reaches the satisfaction threshold and the self-service resolution rate does not reach the resolution rate threshold, the first manual channel is determined as the target channel; If the manual satisfaction rate does not reach the satisfaction rate threshold, and the self-service resolution rate reaches the resolution rate threshold, the self-service channel is determined as the target channel; If the manual satisfaction rate reaches the satisfaction threshold, and the self-service resolution rate reaches the resolution rate threshold, then one of the first manual channel and the self-service channel is selected as the target channel; The user question is assigned to the target channel.

2. The method according to claim 1, wherein: The selecting one of the first manual channel and the self-service channel as the target channel includes: Obtaining idleness information of the first manual channel; According to the idleness information, one of the first manual channel and the self-service channel is selected as the target channel.

3. The method according to claim 1, wherein: The selecting one of the first manual channel and the self-service channel as the target channel includes: Determine whether the current time is in the set busy time period; If the determination result is yes, the self-service channel is selected as the target channel.

4. The method according to claim 1, further comprising: If the manual satisfaction rate does not reach the satisfaction threshold, and the self-service resolution rate does not reach the resolution rate threshold, the second manual channel is determined as the target channel.

5. The method according to any one of claims 1 to 4, further comprising: Obtaining behavior information and / or factor information related to the user problem; The behavior information includes first behavior data generated by the user in the target application to which the user question belongs, and / or second behavior data generated by the target application for the user; the factor information includes a user factor of the user, a business factor related to the business to which the user question belongs, and / or an experience factor of the user under the user question; as well as The method of using the pre-trained manual satisfaction prediction model to predict the manual satisfaction of the user with the manual service after entering the first manual channel according to the user question includes: Using the artificial satisfaction prediction model, predicting the artificial satisfaction according to the user question, and the behavior information and / or factor information; The using of the pre-trained self-service resolution rate prediction model to predict the self-service resolution rate of the user problem after the user enters the self-service channel according to the user problem includes: The self-service resolution rate prediction model is used to predict the self-service resolution rate according to the user problem, the behavior information and / or the factor information.

6. The method according to claim 5, wherein: The artificial satisfaction prediction model includes a plurality of satisfaction prediction models; and The using the artificial satisfaction prediction model to predict the artificial satisfaction according to the user question and the behavior information and / or factor information includes: Using the multiple satisfaction prediction models, predicting multiple satisfaction sub-results according to the user questions, and the behavior information and / or factor information; The manual satisfaction level is determined based on the multiple satisfaction level sub-results.

7. The method according to claim 6, wherein: The multiple satisfaction prediction models include a first hybrid network, and at least one of the following: a first Bert model, a first XGBoost model.

8. The method according to claim 7, wherein: The using of the multiple satisfaction prediction models to predict multiple satisfaction sub-results according to the user questions, and the behavior information and / or factor information includes: The first hybrid network is used to predict a first satisfaction sub-result according to the user question and the behavior information and / or factor information.

9. The method according to claim 8, wherein: The plurality of satisfaction prediction models include the first Bert model; and The method of using the multiple satisfaction prediction models to predict multiple satisfaction sub-results according to the user questions, the behavior information and / or the factor information further includes: The first Bert model is used to predict a second satisfaction sub-result based on the user question.

10. The method according to claim 8 or 9, wherein: The multiple satisfaction prediction models include the first XGBoost model; and The method of using the multiple satisfaction prediction models to predict multiple satisfaction sub-results according to the user questions, the behavior information and / or the factor information further includes: The first XGBoost model is used to predict a third satisfaction sub-result based on the factor information.

11. The method according to claim 5, wherein: The self-service resolution rate prediction model includes a plurality of resolution rate prediction models; and The using the self-service resolution rate prediction model to predict the self-service resolution rate according to the user problem and the behavior information and / or factor information includes: Using the multiple solution rate prediction models, predicting multiple solution rate sub-results according to the user questions, and the behavior information and / or factor information; The self-service resolution rate is determined according to the multiple resolution rate sub-results.

12. The method according to claim 11, wherein the plurality of solution rate prediction models include a second hybrid network and at least one of the following: a second Bert model, a second XGBoost model.

13. The method according to claim 12, wherein: The using of the multiple solution rate prediction models to predict multiple solution rate sub-results according to the user problem, and the behavior information and / or factor information includes: The second hybrid network is used to predict a first resolution rate sub-result based on the user problem and the behavior information and / or factor information.

14. The method according to claim 13, wherein: The plurality of resolution rate prediction models include the second Bert model; and The method of using the multiple solution rate prediction models to predict multiple solution rate sub-results according to the user questions and the behavior information and / or factor information further includes: The second Bert model is used to predict a second resolution rate sub-result based on the user question.

15. The method according to claim 13 or 14, wherein: The plurality of solution rate prediction models include the second XGBoost model; and The method of using the multiple solution rate prediction models to predict multiple solution rate sub-results according to the user questions and the behavior information and / or factor information further includes: The second XGBoost model is used to predict a third solution rate sub-result based on the factor information.

16. The method according to claim 7, wherein: The first hybrid network is obtained by training the initial hybrid network using the following first training step: Acquire multiple first training samples, wherein any first training sample includes a first historical user question, an artificial satisfaction label associated with the first historical user question, and at least one of the following: behavior information and factor information; For a first training sample among the multiple first training samples, the first historical user question in the first training sample, and the behavior information and / or factor information associated with the first historical user question are taken as input, and the artificial satisfaction label associated with the first historical user question is taken as a training label, and the initial hybrid network is trained to obtain a first hybrid network.

17. The method according to claim 16, wherein: The first BERT model is obtained by training the initial BERT model using the following second training step: For a first training sample among the multiple first training samples, the first historical user question in the first training sample is used as input, and the artificial satisfaction label associated with the first historical user question is used as a training label to train the initial Bert model to obtain a first Bert model.

18. The method according to claim 16 or 17, wherein: The arbitrary first training sample includes factor information, and the first XGBoost model is obtained by training the initial XGBoost model using the following third training step: For a first training sample among the multiple first training samples, factor information associated with a first historical user question in the first training sample is used as input, and an artificial satisfaction label associated with the first historical user question is used as a training label to train the initial XGBoost model to obtain a first XGBoost model.

19. The method according to claim 12, wherein: The second hybrid network is obtained by training the initial hybrid network using the following fourth training step: Acquire multiple second training samples, wherein any second training sample includes a second historical user question, a self-service resolution rate label associated with the second historical user question, and at least one of the following: behavior information and factor information; For a second training sample among the multiple second training samples, the second historical user question in the second training sample, and the behavior information and / or factor information associated with the second historical user question are taken as input, and the self-service resolution rate label associated with the second historical user question is taken as a training label, and the initial hybrid network is trained to obtain a second hybrid network.

20. The method according to claim 19, wherein: The second BERT model is obtained by training the initial BERT model using the following fifth training step: For a second training sample among the multiple second training samples, the second historical user question in the second training sample is used as input, and the self-service resolution rate label associated with the second historical user question is used as a training label to train the initial Bert model to obtain a second Bert model.

21. The method according to claim 19 or 20, wherein: The arbitrary second training sample includes factor information, and the second XGBoost model is obtained by training the initial XGBoost model using the following sixth training step: For a second training sample among the multiple second training samples, factor information associated with the second historical user question in the second training sample is used as input, and a self-service resolution rate label associated with the second historical user question is used as a training label to train the initial XGBoost model to obtain a second XGBoost model.

22. A device for channel allocation according to user problems, comprising: An acquisition unit, configured to acquire a user question asked by a user; A manual satisfaction prediction unit is configured to use a pre-trained manual satisfaction prediction model to predict the manual satisfaction of the user to the manual service after entering the first manual channel according to the user question; a self-service resolution rate prediction unit configured to use a pre-trained self-service resolution rate prediction model to predict, according to the user problem, a self-service resolution rate of the user problem after the user enters the self-service channel; The channel allocation unit is configured to, if the manual satisfaction reaches a satisfaction threshold and the self-service resolution rate does not reach a resolution rate threshold, determine the first manual channel as a target channel; if the manual satisfaction does not reach a satisfaction threshold and the self-service resolution rate reaches a resolution rate threshold, determine the self-service channel as the target channel; If the manual satisfaction rate reaches a satisfaction threshold and the self-service resolution rate reaches a resolution rate threshold, one of the first manual channel and the self-service channel is selected as the target channel; and the user problem is assigned to the target channel.

23. A computer-readable storage medium having a computer program stored thereon, wherein: When the computer program is executed in a computer, the computer is caused to execute the method according to any one of claims 1 to 21.

24. A computing device comprising a memory and a processor, wherein: The memory stores executable codes, and when the processor executes the executable codes, the method according to any one of claims 1 to 21 is implemented.

Images