CN115147130A

CN115147130A - Problem prediction method, apparatus, storage medium, and program product

Info

Publication number: CN115147130A
Application number: CN202210771557.XA
Authority: CN
Inventors: 陈志钊
Original assignee: Alibaba China Co Ltd
Current assignee: Alibaba China Co Ltd
Priority date: 2022-07-01
Filing date: 2022-07-01
Publication date: 2022-10-04

Abstract

The present application provides a problem prediction method, apparatus, storage medium and program product, the method comprising: acquiring an interactive behavior sequence of a user within current preset time; wherein the interaction behavior sequence comprises a browsed page sequence; predicting the type of the problem encountered by the user based on the problem classification model according to the interaction behavior sequence, and predicting the problem encountered by the user based on the problem prediction model corresponding to the type; and determining a solution pushed to the user according to the predicted problem. According to the method and the device, the page sequence corresponding to the browsing application program of the user in the preset time is classified, the type corresponding to the page sequence is determined, then the model matched with the type is selected to predict the problem, the problem which possibly exists in the user at present is accurately predicted, the problem of the user is found in time, the accuracy, the real-time performance and the effectiveness of the problem prediction are improved, the solution is recommended to the user, the problem of the user is solved at the first time, and the service experience of the user is improved.

Description

Problem prediction method, apparatus, storage medium, and program product

Technical Field

The present application relates to the field of artificial intelligence technologies, and in particular, to a problem prediction method, device, storage medium, and program product.

Background

With the continuous development of the artificial intelligence technology, the intelligent customer service robot has the capabilities of question answering and conversation of a knowledge base, and can solve the user problems smoothly and accurately.

However, the customer service robot passively receives user consultation, requires a user to find a customer service robot entrance and express a service demand, wastes time and labor, wastes resources, and is difficult to find user problems timely and effectively, and solves the user problems at the first time, so that the user experience is poor.

Disclosure of Invention

The embodiments of the present application mainly aim to provide a problem prediction method, device, storage medium, and program product, so as to perform real-time problem and intention prediction, thereby improving the service experience of a user.

In a first aspect, an embodiment of the present application provides a problem prediction method, including:

acquiring an interactive behavior sequence of a user within current preset time; wherein the sequence of interaction behaviors comprises a sequence of browsed pages;

predicting the types of the problems encountered by the users based on a problem classification model according to the interaction behavior sequence, and predicting the problems encountered by the users based on a problem prediction model corresponding to the types;

Determining a solution to push to the user based on the predicted problem.

Optionally, predicting the type of the problem encountered by the user based on a problem classification model according to the interaction behavior sequence, and predicting the problem encountered by the user based on a problem prediction model corresponding to the type, including:

inputting the feature vector corresponding to the interaction behavior sequence into the problem classification model, and determining the type of the problem encountered by the user;

if the type is an operation and sales class or an account class, processing the interaction behavior sequence of the user based on a sequence model to obtain a corresponding problem;

and if the type is a transaction type, processing the interaction behavior sequence of the user and the transaction information corresponding to the user based on a deep interest evolution network model to obtain a corresponding problem.

Optionally, the method further includes:

acquiring historical conversation data of historical users and intelligent customer service, and extracting corresponding problems according to the historical conversation data;

according to the historical interaction behavior sequence of the historical user before the historical conversation data and the extracted problems, constructing a training data set corresponding to a sequence model, and training the sequence model; and/or the presence of a gas in the gas,

And constructing a training data set corresponding to the deep interest evolution network model according to the historical interaction behavior sequence of the historical user before the historical conversation data, the transaction information and the extracted problems, and training the deep interest evolution network model.

Optionally, the sequence model is a natural language model for processing a sequence; training the sequence model, including:

performing word segmentation operation on the historical interactive behavior sequence in the training data set and the extracted problems to obtain a word segmentation sequence;

inputting the word segmentation sequence into the natural language model, and pre-training the natural language model by taking the predicted next sentence as a training target;

and taking the extracted problem as a supervision signal, and carrying out fine tuning training on the pre-trained natural language model according to the historical interactive behavior sequence.

Optionally, the obtaining an interaction behavior sequence of the user within the current preset time includes:

generating streaming data containing an interaction behavior sequence according to an interaction log of a user in the current preset time through a stream computing frame;

and screening the streaming data based on a rule logic library, and putting the screened streaming data into a message queue, wherein the streaming data in the message queue is used for processing the problem classification model.

Optionally, the screening the streaming data based on the rule logic library, and placing the screened streaming data into a message queue, includes:

screening the streaming data according to a dictionary and rules in the rule logic library;

the dictionary comprises a page dictionary and/or a page keyword dictionary;

the dictionary is determined through a history page browsed by a history user and a corresponding problem;

the rule is used for expressing whether the dictionary is used as a black list or a white list, and/or the screening operation of the page sequence is completed when how many pages in the page sequence browsed by the user are matched with the dictionary.

Optionally, the method further includes:

acquiring historical conversation data of a plurality of historical users and intelligent customer service, and extracting corresponding problems according to the historical conversation data to obtain a problem set;

determining a page set browsed by the plurality of historical users and a page keyword set according to a historical browsing page sequence of the plurality of historical users before the historical conversation data;

for any problem in a problem set, calculating mutual information between each page in the page set and the problem, and mutual information between each page keyword in the page keyword set and the problem, and selecting a target page and a target page keyword corresponding to the problem from the page set and the page keyword set by the calculated mutual information;

And fusing target pages corresponding to the problems in the problem set to obtain a page dictionary, and fusing target page keywords corresponding to the problems to obtain a page keyword dictionary.

In a second aspect, an embodiment of the present application further provides a problem prediction method, including:

acquiring an interaction behavior sequence of at least one buyer corresponding to the shop within current preset time; wherein the sequence of interaction behaviors comprises a sequence of browsed pages;

predicting the types of the problems encountered by the buyers based on a problem classification model according to the interaction behavior sequence, and predicting the problems encountered by the buyers based on a problem prediction model corresponding to the type;

and determining a solution pushed to a seller corresponding to the shop according to the predicted problem.

In a third aspect, an embodiment of the present application provides an electronic device, including:

at least one processor; and

a memory communicatively coupled to the at least one processor;

wherein the memory stores instructions executable by the at least one processor to cause the electronic device to perform the method of any of the above aspects.

In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, in which computer-executable instructions are stored, and when a processor executes the computer-executable instructions, the method according to any one of the above aspects is implemented.

In a fifth aspect, the present application provides a computer program product, which includes a computer program that, when executed by a processor, implements the method of any one of the above aspects.

According to the problem prediction method, the problem prediction device, the problem prediction storage medium and the program product, the interaction behavior sequence of the user in the current preset time can be obtained; wherein the interaction behavior sequence comprises a browsed page sequence; further, according to the interaction behavior sequence, predicting the type of the problem encountered by the user based on the problem classification model, and predicting the problem encountered by the user based on the problem prediction model corresponding to the type; further, a solution to be pushed to the user is determined according to the predicted problem. The method and the device can classify the page sequence corresponding to the browsing application program of the user within the preset time, determine the type corresponding to the page sequence, further select the model matched with the type to predict the problem, accurately predict the problem possibly existing in the user at present, find the problem of the user in time, improve the accuracy, the real-time performance and the effectiveness of problem prediction, recommend a solution to the user, solve the problem of the user at the first time, and further improve the service experience of the user.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the application and, together with the description, serve to explain the principles of the application.

Fig. 1 is a schematic view of an application scenario according to an embodiment of the present application;

fig. 2 is an interface display schematic diagram of a terminal device according to an embodiment of the present disclosure;

fig. 3 is a schematic flowchart of a problem prediction method according to an embodiment of the present application;

fig. 4 is a schematic structural diagram of a flow calculation framework according to an embodiment of the present disclosure;

FIG. 5 is a schematic diagram of a process for determining a problem classification according to an embodiment of the present application;

FIG. 6 is a schematic diagram illustrating training of a deep interest evolution network model according to an embodiment of the present disclosure;

FIG. 7 is a diagram illustrating pre-training of a natural language model according to an embodiment of the present disclosure;

fig. 8 is a flowchart of determining a dictionary based on mutual information according to an embodiment of the present application;

FIG. 9 is a flowchart illustrating an online calculation of a problem prediction method according to an embodiment of the present disclosure;

FIG. 10 is a schematic flow chart illustrating another problem prediction method according to an embodiment of the present application;

fig. 11 is a schematic structural diagram of a problem prediction apparatus according to an embodiment of the present application;

Fig. 12 is a schematic structural diagram of another problem prediction apparatus provided in an embodiment of the present application;

fig. 13 is a schematic structural diagram of an electronic device according to an embodiment of the present application.

With the above figures, there are shown specific embodiments of the present application, which will be described in more detail below. These drawings and written description are not intended to limit the scope of the inventive concepts in any manner, but rather to illustrate the inventive concepts to those skilled in the art by reference to specific embodiments.

Detailed Description

Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. The following description refers to the accompanying drawings in which the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application.

The terms referred to in the present application are explained first:

and (3) flow calculation: the method refers to a computing architecture, which can acquire mass data from different data sources in real time, and obtain valuable information through real-time analysis and processing, so as to analyze large-scale flow data in a constantly changing motion process in real time, capture possibly useful information, and send the result to the next computing node.

A stream calculation engine: refers to a real-time computing engine developed based on an open-source distributed computing framework for implementing stream computing.

The recommendation system comprises: it refers to using an e-commerce website to provide commodity information and suggestions to a customer, help the user decide what product should be purchased, simulate a salesperson to help the customer complete the purchasing process, such as providing a sorted recommendation list of personalized items (item) to the user according to the user's historical preferences and constraints, i.e. recommendation results can be generated according to user preferences, commodity characteristics, user-commodity transactions and other environmental factors (such as time, season, location, etc.), and the recommended items can include movies, books, restaurants, news items, etc. in this application, the personalized items can be questions in the customer service system, and can also be referred to as intentions or scenes.

C, end user: in the context of an e-market, it is often referred to as a consumer or buyer.

Through the development in recent years, the intelligent customer service robot has knowledge base question-answering and conversation capabilities, can solve user problems more smoothly and accurately, and meanwhile, the intelligent customer service robot with better experience can predict the user problems based on the user problems, so that the associated guess of the user problems is realized.

However, various operations leave behavior traces during the application program usage process of the user, and therefore, in some technologies, different messages may be actively recommended to different users by extracting content that may be of interest to the user based on user portrayal or a user log generated when the user uses the application program. For example, a crowd-label-based push system can actively push different messages, promote a promotional document, and the like to different crowds.

For example, the pushing system may perform circle of people meeting a set condition by running a Structured Query Language (SQL) in a background data table, or manually screen the target user through a combination of crowd tags based on a tag system of a demographic and an interest model, so as to push a message.

However, the active message pushing mode of the crowd-based pushing system needs to mark the user by relying on the crowd labels and attributes, and the crowd label system needs to be developed in a long period, and crowd labels supported by a third party are often incompatible and have inconsistent calibers; on the other hand, the tag system based on the demographic and interest model is a relatively long-term attribute tag, and cannot be changed greatly in a short term, so that the accuracy and the real-time performance of pushing the message to the user in a tag printing mode are poor.

It should be noted that, the push system mostly performs one-time push for campaign promotion and promotion, is not an instant push scenario customized for the intelligent customer service robot, is difficult to adapt to an instant "answer or solve problem" scenario, has poor push timeliness, and the message push timing is not the first time when a problem occurs, so that after a target is pushed in the manner described above, a certain hysteresis often exists, and the intention of a user to open a push message is greatly reduced.

However, since the robot has the capability of problem prediction, in the intelligent service, the problem of the user can be actively predicted by acquiring the user log in real time, and the user is prompted to click the push message to help solve the problem by sending the push message to the user at the first time.

Therefore, for the intelligent customer service robot, the problems possibly encountered by the user can be predicted in real time by using the flow calculation capacity, and related solutions can be pushed, so that the quality and the temperature of the service are improved. For example, in the process of using an electronic product by a C-end user, the system records a Uniform Resource Locator (URL) of the user accessing a browsing page, the URL of each page corresponds to an operation of the user, and a behavior track of the user can be formed by temporally sequencing the URLs, where the behavior track includes an instant intention and a demand of the user, and further, based on the behavior track of the user and monitoring and predicting abnormal behaviors, problems that the user may encounter can be predicted in real time, so as to recommend a relevant solution.

In view of this, an embodiment of the present application provides a problem prediction method, which may process log data of a user browsing a page in an application program to a message trigger source converted into stream computing, analyze a type of a problem that the user may encounter in the application program by using an algorithm model, and select a corresponding model based on the type of the problem to perform real-time problem and intention prediction, where fig. 1 is an exemplary application scenario diagram related to the embodiment of the present application. As shown in fig. 1, the problem prediction method provided in the embodiment of the present application may be applied to an application scenario shown in fig. 1, where the application scenario includes: a terminal device 101, an analysis platform 102 and a processing platform 103; specifically, the analysis platform 102 may obtain a page sequence of an application program browsed by the terminal device 101 in 30 minutes, further determine a type of the page sequence by using a problem classification model based on the obtained page sequence browsed by the terminal device 101 in 30 minutes, select a problem prediction model corresponding to the type, such as problem prediction model 1, according to the type corresponding to the page sequence, predict a problem 1 that may be encountered, and send the problem 1 to the processing platform 103, further, the processing platform 103 determines a solution 1 to be pushed to a user according to the problem 1, and sends the determined solution 1 to the terminal device 101 for display for the user to view.

It can be understood that, at preset intervals, the processing platform 103 may adjust the solution corresponding to the problem 1 based on the feedback information of the user, so that the solution better meets the requirements of the user, and the user experience is improved.

The analysis platform 102 includes massive data, and may classify the data, and further determine and store a problem prediction model corresponding to each type for problem prediction, so that the analysis platform 102 may be applicable to a variety of scenarios, taking the analysis platform 102 as an e-commerce platform as an example, the analysis platform may be divided according to the type of the e-commerce platform, each type corresponds to a problem prediction model, if 3 types exist, corresponds to the problem prediction model 1, the problem prediction model 2, and the problem prediction model 3, and each prediction model may predict one or more problems in the corresponding type.

It should be noted that, the determining of the type of the page sequence by using the problem classification model, invoking the problem prediction model corresponding to the type to predict the problem encountered by the user, and determining the execution step of the solution pushed to the user according to the problem are merely examples, and may also be performed by only one platform, such as a cloud, or may be performed by splitting more platforms and more modules, which is not specifically limited in the embodiment of the present application.

Optionally, a log corresponding to a page of an application browsed by the terminal device 101 within 30 minutes may be obtained in the following manner, fig. 2 is an interface display schematic diagram of the terminal device provided in the embodiment of the present application, as shown in fig. 2, the terminal device may obtain browsing log data according to an application interface opened by a user and a touch operation on the application interface, for example, a user clicks a "my" button in an opened electronic commerce interface through a touch operation, jumps to an individual user interface, clicks a "refund/after-sale" button in my order through a touch operation again, browses refund/after-sale related data, and further, the terminal device obtains browsing log data.

Further, after acquiring the browsing log data, the terminal device may send the browsing log data to a log storage module, where the log storage module may be any device or platform capable of storing log data, such as a log service data center or a data warehouse. According to the scheme provided by the application, the log data can be converted into the streaming data in real time by using modes such as stream calculation and the like, and a downstream active service process is started. The log service data center or the data warehouse stores persistently, and the second-level delay can be achieved by the flow calculation triggering in the scheme.

Optionally, the active service process may determine a problem that the user may encounter, such as "the user needs to know a refund or an after-sales processing process", and further push a corresponding solution, such as "a refund process/an after-sales process", to the user, so as to help the user solve the problem that the user may encounter in time.

Therefore, the problem prediction method provided by the embodiment of the application can actively predict the problems which may exist in the user without asking questions of the customer service robot by the user, can find the problems in time while saving time, and can determine the types of the problems which may be encountered by the user by using the algorithm model when predicting the problems which may be encountered by the user, so as to select the corresponding model prediction problems based on the types, improve the accuracy and real-time performance of problem prediction, and can actively push the corresponding solutions to the user, thereby realizing accompanying services and improving the service experience of the user.

The technical means of the present application will be described in detail with specific examples. The following several specific embodiments may be combined with each other, and details of the same or similar concepts or processes may not be repeated in some embodiments. Embodiments of the present application will be described below with reference to the accompanying drawings.

Exemplarily, fig. 3 is a schematic flowchart of a problem prediction method provided in an embodiment of the present application. The embodiment can be applied to any device capable of realizing data processing, such as a server and the like. As shown in fig. 3, the method may include:

s301, acquiring an interactive behavior sequence of a user within current preset time; wherein the sequence of interaction behaviors comprises a sequence of browsed pages.

In the embodiment of the application, the interactive behavior sequence is used for embodying the interactive behavior of a user and an application program in terminal equipment, and comprises a browsed page sequence, the page sequence comprises a sequence formed by information of a plurality of pages browsed by the user in current preset time, and the information of the pages comprises page names and/or page URLs; the page names may include, for example, order detail pages, help, item detail pages, and the like.

Optionally, in this step, the interaction behavior sequence of the user within the current preset time may be obtained in real time. The preset time refers to time set by a preset system for processing the collected data in real time, and the preset system can be a system for processing the data set by the cloud, so that the browsing log data of the user can be processed immediately, and the accuracy of prediction is further improved, for example, the preset time can be 30 minutes.

The interaction behavior sequence may be a page sequence that is obtained within a current preset time of a user and meets preset requirements, where the preset requirements may be a preset number, a specified page, or other requirements, and this is not specifically limited in the embodiment of the present application.

For example, in the application scenario of fig. 1, a page sequence, that is, 30 URL records, corresponding to 30 real-time shopping pages viewed by the terminal device 101 in a certain application program in the current 30 minutes may be obtained.

S302, predicting the types of the problems encountered by the users based on the problem classification models according to the interaction behavior sequences, and predicting the problems encountered by the users based on the problem prediction models corresponding to the types.

In the embodiment of the present application, the types of the problems may be classified according to the nature of the input data, and if the input data is only an interactive behavior sequence, a model that only processes the sequence may be used, and if the input data is an interactive behavior sequence and other information, such as an order state, a model that can process the sequence and other information at the same time needs to be used.

Specifically, the types include at least the following two types: predicting a problem based only on the sequence of the interactive behaviors, predicting a problem based on the sequence of the interactive behaviors, and other information; and calling a corresponding model to predict the problems encountered by the user according to the type. The questions may include a right-maintaining question, a product recommendation, a logistics question, store evaluation information, and the like, which is not specifically limited in the embodiment of the present application.

For example, the method and the device for predicting the user problems can call the corresponding model based on the interactive behavior sequence to predict the user problems, for example, the method and the device for predicting the user problems call the corresponding model based on the page sequence corresponding to the application program interface browsed by the user in the preset time period, or call the corresponding model based on the interactive behavior sequence and other information to predict the user problems; the other information may include an order state, a logistics state, an optional problem scene, and the like, which is not specifically limited in this embodiment of the present application, and the more the types and contents of the other information include, the higher the accuracy of predicting the problem is, for example, according to the input of a model that a user browses page browsing data, the order state, the logistics state, and the optional problem scene corresponding to an application program interface within a preset time period, the corresponding model is further invoked to obtain the probability of selecting each problem by the user, that is, the problem encountered by the user is predicted.

Optionally, the output data may be classified according to the nature of the output data, for example, the output data with little influence on the user, such as data related to user preference, may use a model with general accuracy but high speed, while the output data with large influence on the user, such as data related to transaction, may use a model with higher accuracy, or may be classified according to the amount of calculation to obtain the output data, for example, the output data has large amount of calculation, a model with higher calculation accuracy may be used, otherwise, a model with general calculation accuracy may be used, which may reduce the burden of the preset system.

It should be noted that, in the embodiment of the present application, specific models corresponding to the problem classification model and the problem prediction model are not limited, and the problem classification model and the problem prediction model may satisfy the following conditions: after the interactive behavior sequences are classified by using the problem classification model, a problem prediction model corresponding to the type can be called to predict problems possibly encountered by the user.

S303, determining a solution pushed to the user according to the predicted problem.

Specifically, according to the predicted problem, the solution of the user may be obtained by matching the corresponding message template, and the solution is pushed to the user.

In this step, the solution may refer to a solution actively pushed to the user for the predicted problems that the user may encounter, the solution may be stored in a lookup table in advance, and directly called when used, each problem has a corresponding solution, and the corresponding solution may be searched in the lookup table by the id corresponding to the problem.

Optionally, the solution may be a document, a link, a card, or the like, and may be pushed to the user in a manner of a message, a short message, a telephone, or the like in the application program, for example, the link of the refund may be recommended to the user in a form of a message.

Illustratively, taking a travel platform as an example, browsing page information of an application program corresponding to the travel platform by a user, and if a problem encountered by the user is predicted to be a right-maintaining problem based on a corresponding problem prediction model, actively recommending a right-maintaining flow to a display interface of terminal equipment of the user; if the problem encountered by the user is predicted to be a travel route selection problem based on the corresponding problem prediction model, actively recommending a travel route, a travel agency, a travel strategy and the like in the season to a display interface of the terminal equipment of the user; if the problem encountered by the user is predicted to be a ticket business problem based on the corresponding problem prediction model, actively recommending a ticket buying flow, a ticket changing flow or a ticket returning flow to a display interface of the terminal equipment of the user.

Taking a takeaway platform as an example, browsing webpage information of an application program corresponding to the takeaway platform by a user, and actively pushing takeaway related information which is possibly selected to a display interface of terminal equipment of the user if the problem encountered by the user is predicted to be a rice ordering problem based on a corresponding problem prediction model; taking an e-commerce platform as an example, by counting webpage information of an application program corresponding to the e-commerce platform browsed by a user, real-time commodity comment, service data and the like, if the problem encountered by the user is predicted to be a shopping problem based on a corresponding problem prediction model, relevant information of possibly selected shop merchants is actively pushed to a display interface of terminal equipment of the user.

Therefore, the method and the device can obtain the page sequence browsed by the user within the current preset time; predicting the type of the problem encountered by the user based on the problem classification model, and actively predicting the problem encountered by the user based on the problem prediction model corresponding to the type; further, a solution to be pushed to the user is determined according to the predicted problem. The method and the device can classify the page sequence corresponding to the browsing application program of the user within the preset time, determine the type corresponding to the page sequence, further select the model matched with the type to predict the problem, accurately predict the problem possibly existing in the user at present, find the problem of the user in time, improve the accuracy, the real-time performance and the effectiveness of problem prediction, recommend a solution to the user, solve the problem of the user at the first time, and further improve the service experience of the user.

In view of the above considerations, the problem prediction method provided by the present application may be implemented based on a stream computing framework, and fig. 4 is a schematic structural diagram of the stream computing framework provided by the embodiment of the present application, as shown in fig. 4, the stream computing framework is a layered user intention prediction framework, and may predict a problem that may be encountered by a user based on browsing log data of the user, so as to improve accuracy of problem prediction. Specifically, the flow calculation framework comprises two parts of components, namely a flow calculation message trigger source consisting of a flow calculation self-defined function module, an algorithm model Http online service module (algorithm service for short) and a rule logic library, and a message consumption prediction component consisting of a recall module (Matching), a sorting module (Ranking) and modules for fatigue control and push control; the stream computing message trigger source integrates middleware such as a stream computing engine and a message middleware (MetaQ), and message consumption prediction is used for calling various algorithm models in a layering mode, such as calling algorithm models in a recall algorithm model library and a sequencing algorithm model library to predict problems possibly encountered by a user.

Illustratively, a user log of a user browsing page is processed into streaming data of a key value pair through a stream calculation engine, further, a stream calculation self-defined function module calls an algorithm in an algorithm model Http online service module, a large amount of streaming data can be screened through a key page/key word calculated based on mutual information and a rule logic library, and the obtained screened streaming data is put into a message middleware as a message for message consumption prediction.

Furthermore, in the process of predicting message consumption, the messages sequentially pass through the recall module and the sequencing module to predict problems possibly encountered by the user, and then are fed back to the offline data computing (scheduling) service based on the push control module, the feedback content further comprises a log table and a push table, the log table is used for storing user logs, the push table is used for updating mutual information, the mutual information is used for measuring the information amount contributed by the occurrence of a certain keyword/key page to the occurrence of a knowledge point, and the knowledge point is a problem encountered by the user, so that a closed loop is formed.

The purpose of current limiting is to reduce accumulation of messages, when the message amount is larger than a threshold value, part of messages can be discarded, and the fatigue control module is used for setting the pushing times to a user, so that the phenomenon that the user feels dislike due to too many pushing times is reduced.

In the embodiment of the present application, the feature vector may refer to a two-dimensional vector obtained by vectorization, and may also be referred to as an embedded expression vector. Optionally, the interaction behavior sequence may be vectorized using the word2vec model. Illustratively, a user browsing log may be obtained in real time by accessing an application program, and a serialized page browsing path (i.e., an interaction behavior sequence) of the user, such as page 1 to page N, is determined. In an Embedding presentation (Embedding presentation) stage, a serialized page browsing path of a user is coded into a two-dimensional vector, and then problem classification is carried out according to the two-dimensional vector.

In this step, the problem classification model is used for classifying the problems encountered by the user, predicting whether the current user has one or more of the three types of problems of account, transaction and marketing, and labeling the problems, namely labeling the problems of the user so as to call a corresponding problem prediction model to predict the problems of the user based on the interactive behavior sequence; the problem prediction model includes: problem prediction models based on language models, such as sequence models and problem prediction models based on Deep learning point selection (CTR) models, such as Deep Interest Evolution Network (DIEN) models, wherein the sequence models are used for solving problem prediction under marketing and account class labels; the DIEN model is used for solving problem prediction under a transaction label, namely performing secondary sequencing on problems encountered by a user.

The CTR model is introduced in the embodiment of the application, so that the problem prediction problem is converted into a matching problem for judging a User (User) and a problem (item), namely, the problem that the current User can click (1) or cannot click (0) is judged, and then the problem is solved through 0/1 scoring, specifically, the feature vector input by the CTR model comprises a User page browsing behavior sequence, discrete features and continuous features; the discrete characteristics such as orders, logistics, service states, problem ids and the like, the continuous characteristics such as order prices, click rates of articles and the like, and the characteristic vectors are input into the CTR model, so that the output 'predicted click rate' of the model can be obtained, the value is 0-1, namely whether a user clicks the article, the click is 1, the click is 0, and further, the problem with the corresponding high predicted value in the problem can be selected as the problem which is preferentially pushed by the user.

In view of different data forms of different problem labels, different problem classification models are used for prediction, if the type is a marketing type or an account type, problems encountered by the user are predicted based on a sequence model, the sequence model may be a natural language model or any other model capable of processing a sequence, and this is not specifically limited in the embodiment of the present application, for example, if the sequence model is a Bert-MM-1 model, and if the type is a transaction type, problems encountered by the user are predicted based on a deep interest evolution network model.

Predicting the problem type to be an operation and marketing type or an account type only based on the interaction behavior sequence, and predicting the problem type to be a transaction type based on the interaction behavior sequence and other information; the marketing category is the problems encountered by the user for purchasing commodities, including the problems of playing methods, how to get red packages, coupons and the like; the account type is the problem encountered by the user for modifying personal information, and comprises the problems of modifying a receiving address, modifying an account binding mobile phone number and the like; the transaction category is a problem encountered by a user in transaction, and includes a problem of how to apply for a refund, a problem of how to query a logistics state, a problem of how to query an order state, and the like.

Specifically, the problem classification model converts the interactive behavior sequence into a two-dimensional vector of a page browsing path code by vectorizing the interactive behavior sequence, inputs the two-dimensional vector into a corresponding Convolutional Neural Network (CNN), extracts a serialized hidden mode through the CNN, and outputs a classification corresponding to a problem as a prediction result; further, based on the type corresponding to the problem, a corresponding problem prediction model is selected, and the interaction behavior sequence is used as the input of the model, so that the output result of the probability of each problem selected by the user is obtained. For example, fig. 5 is a schematic diagram of a process for determining a problem classification according to an embodiment of the present application, and as shown in fig. 5, an implicit mode of a serialized embedded expression vector may be extracted through CNN, that is, a problem classification is obtained as a prediction result through processing of convolution, max pooling, full-connected neural network and a classifier.

Optionally, a process of vectorizing the interactive behavior sequence may be unsupervised, specifically, in a process of training the coding model, based on a page browsing path of a large number of users, unsupervised "semantic information" is mined and stored as the coding model, and the coding model may be understood as a word2vec model, that is, the page browsing path and a corresponding two-dimensional vector are input into the word2vec model for training, so as to obtain the trained word2vec model.

Therefore, due to the fact that the data forms corresponding to the interactive behavior sequences are different, the problem prediction method and the problem prediction device can call the corresponding problem prediction model based on the types of the problems encountered by the user and perform problem prediction by means of the interactive behavior sequences, have corresponding processing modes aiming at different data types, enable different data to predict the problems of the user, and improve flexibility of a problem prediction scene.

Optionally, historical session data of historical users and intelligent customer service can be acquired, and corresponding problems are extracted according to the historical session data;

In the embodiment of the application, the historical user may refer to a user who has a browsing record in an application program and has actively consulted a problem within a certain period of time, for example, the user 1 who has browsed a refund page of the application program and has consulted a refund problem, the historical session data may refer to session data of problem communication between the historical user and an intelligent customer service, for example, a series of session contents of a refund process of an article having consulted the customer service by the user 1, and a corresponding problem may be extracted based on a natural language processing model according to the historical session data.

Specifically, according to the embodiment of the application, a corresponding training data set is constructed through a historical interaction behavior sequence of a historical user before historical session data and an extracted problem, wherein the extracted problem is a real result, namely a label, corresponding to the training data and is used for improving the training accuracy.

Illustratively, the sequence model can be any model that can process sequence data, such as a natural language model. The training set for constructing the sequence model can be obtained by the following method: the method comprises the steps of obtaining an interactive behavior sequence of at least one user in a certain application program interface within a period of time as characteristic data, extracting corresponding problems as labels corresponding to the characteristic data based on session data of each user and intelligent customer service within the period of time, wherein the characteristic data and the labels form a training set of a sequence model, and further training the sequence model by using the training set, so that the output of the sequence model can be close to the extracted corresponding problems, and the trained sequence model is obtained.

For most trading scenes, some 'characteristic factors' such as order states and logistics states need to be considered in model training, and sequence models become unavailable, so that one of deep learning CTR models such as a DIEN model is adopted for trading classes, namely, the DIEN model is supervised trained by using labeled data and is used for predicting user problems in combination with information such as discrete orders under the large scene of 'trading' after training is finished, and the DIEN model is a prediction model for solving the prediction problem end to end.

Optionally, the transaction information may include: the information of orders, logistics and the like can correspond to discrete characteristics or continuous characteristics, for example, the state information is discrete characteristics, and the price information is continuous characteristics. FIG. 6 is a schematic diagram illustrating training of a deep interest evolution network model according to an embodiment of the present disclosure; as shown in fig. 6, in the training of the DIEN model, a historical interaction behavior sequence is input as one path, price information is input as another path, for example, a discrete feature such as an order state, a logistics state, or a continuous feature such as an order price is input as another path, a corresponding problem scene is input as another path, the input data is input into the DIEN model for training, so that the trained DIEN model can be obtained, an output result of the DIEN model can include a click probability for the problem scene, the probability is used for determining a probability measure of a problem encountered by a user, and an output result of the control model in the first N values with the maximum click probability can improve the accuracy of the problem, wherein the value of the probability measure is between 0 and 1, and N is a positive integer greater than 1.

Therefore, the training data set can be constructed based on the historical conversation data of the intelligent customer service, so that the model can be trained according to the problem of real consultation of the user, and the accuracy of model prediction is improved.

In the embodiment of the application, as the problem prediction of marketing and account classes can reach higher accuracy and precision rate only by inputting pure behavior sequences, the sequence models such as the natural language models can be used for processing the problem prediction of marketing and account classes, but due to the particularity of the URLs, a plurality of meaningless words such as 'www' and 'myPage' exist in the historical interactive behavior sequences, and are considered as a word, so that a plurality of vectors are obtained after the page word segmentation operation is performed on the historical interactive behavior sequences, a training file is further formed, the natural language model is pre-trained by using the training file, and the natural language model can be any model for calculating the probability of the next word segmentation in the sequence data, such as a Bert model.

Specifically, word segmentation is performed on an interactive behavior sequence of a user in a certain time period, namely, word segmentation is performed on a page name and a page URL (uniform resource locator), for example, an item of a certain page is subjected to word segmentation, an original page item is changed into a sequence of a plurality of keywords after word segmentation, namely, a single item is processed into more semantic information based on modeling of the keywords, after word segmentation is performed, the interactive behavior sequence is changed into a plurality of vectors, and a natural language model is pre-trained by using the vectors to obtain an unsupervised natural language model capable of inputting the behavior sequence.

For example, fig. 7 is a schematic diagram illustrating a natural language model provided in an embodiment of the present application being pre-trained; as shown in fig. 7, a training data set is obtained, and a word segmentation operation is performed on a history interaction behavior sequence and an extracted problem in the training data set, the word segmentation operation on the history interaction behavior sequence may be to extract words from a URL, and the word segmentation operation on the extracted problem may be to extract words from a problem corresponding to the URL, where, for example, a first word is extracted from the URL corresponding to "user baseinfoset name app mtp app note website app ikc redirect index newmenter", whether a word is omitted from the problem corresponding to the URL corresponding to the chinese "how account number" corresponds to the word extracted from the URL, so as to obtain a word segmentation sequence, and further, the word segmentation sequence is input into a natural language model to predict a next sentence as a training target, and a natural language model is pre-trained, that is, the input word sequence is processed to obtain a pre-trained natural language model.

In this step, the fine tuning training may refer to finetune training, which is training a new model based on a pre-trained natural language model by using a historical interaction behavior sequence and an extraction problem, and may obtain a model with a better effect by using a smaller number of iterations without completely retraining the natural language model, thereby improving the efficiency of model training.

Therefore, the natural language model can better process the behavior sequence through the pre-training, and when the pre-training is carried out, the interaction sequence and the corresponding problem can be used as model input together, so that the pre-trained model can more accurately extract the meaning of the interaction sequence, and the pre-training effect is improved.

Optionally, the obtaining of the interaction behavior sequence of the user within the current preset time includes:

generating streaming data containing an interaction behavior sequence according to an interaction log of a user in the current preset time through a stream calculation framework;

In the embodiment of the present application, the streaming data is a key-value pair formed by processing a user log through a computing engine in a streaming computing framework, for example, if a user clicks a button in an application program or accesses a page, a piece of log data is correspondingly recorded in a log system, and further, the streaming data is processed into a corresponding key-value pair through the computing engine, that is, a key-value pair is formed by taking an Identity card identification number (ID) of the user as a key and taking a sequence of URLs as a value.

Specifically, when processing log data, the stream computing framework packages implementation logic of the processed log data into a jar packet for processing, generates stream data containing an interactive behavior sequence, further submits the stream data to a stream computing engine for execution in a form of a stream computing custom function, and further calls an algorithm model Http online service interface, wherein the interface is used for screening the stream data, namely judging whether the stream data represents that a user has a consultation requirement based on a dictionary, the dictionary comprises a plurality of keywords and/or key pages, and if the consultation requirement exists, the stream data is written into a message queue (MetaQ) in a message form; if the consultation requirement does not exist, the streaming data is discarded, and after the interactive behavior sequence of the user in the next preset time is obtained, the judgment is carried out again through the mode.

Whether consultation needs exist in the streaming data is judged, that is, whether consultation needs exist is determined by judging whether the number of keywords and/or key pages in a dictionary hit by the streaming data is greater than a preset threshold, the preset threshold is not specifically limited in the embodiment of the present application, for example, the number of keywords in the streaming data is judged to be greater than 5, and/or the number of key pages in the streaming data is judged to be greater than 1.

Optionally, the interactive behavior sequence may further include an interactive commodity sequence, a button event of the user on a page, and the like, for example, the user clicks a button event of a collection operation on a certain page.

Therefore, the streaming data can be generated based on the streaming calculation framework, the streaming data meeting the requirements are screened by using the rule logic base, the streaming data meeting the requirements are screened out, the streaming data meeting the requirements are input into the problem classification model to be processed, interference of useless data is reduced, and the problem classification accuracy is improved.

Optionally, screening the streaming data based on a rule logic library, and placing the screened streaming data into a message queue, including:

the dictionary comprises a page dictionary and/or a page keyword dictionary;

the dictionary is determined through historical pages browsed by historical users and corresponding problems;

In the embodiment of the present application, the dictionary is determined based on mutual information, where the mutual information refers to a useful information measure for expressing the amount of information about a keyword (keyword) included in a problem encountered by a user, that is, the amount of information that a certain keyword contributes to a problem encountered by a user is measured, and may be determined by the following formula:

wherein x is _i Denotes the ith keyword, y, of all possible keywords extracted from the URL _i Represents the ith question of all possible questions encountered by the user, I represents the corresponding quantity, I (Y, X) is used to measure the contribution of the keyword to the question encountered by the user, P (X = X) _i ,Y＝y _i ) Denotes X = X in history data _i And Y = Y _i Probability of simultaneous occurrence, where the historical data is a problem extracted from the URL and corresponding to the URL obtained within a period of time, P (X = X) _i ) Indicating X = X in URL in history data _i Probability of occurrence, P (Y = Y) _i ) Indicating Y = Y in the history data _i The probability of occurrence, which can be obtained by calculating the corresponding data in the history data, such as P (X = X) _i ) The method can be obtained by calculating the ratio of the number of keywords extracted from the URL in the historical data to the number of possible participles formed by all data in the historical data.

The dictionary comprises a page dictionary and/or a page keyword dictionary; the page dictionary comprises at least one page name or URL; the page keyword dictionary includes keywords in a page name or URL.

In the embodiment of the present application, the keyword is split from a page name or a URL, for example, if the page name is a refund detail page, the keyword may be split into three keywords of "refund", "detail", and "page", and further, mutual information of the three keywords is calculated respectively, and a keyword whose mutual information is greater than a set threshold is a page keyword in a dictionary.

If the dictionary only comprises the page dictionary, when the streaming data is screened, determining how many page names or URLs in the streaming data hit the corresponding page names or URLs in the dictionary; if the dictionary only comprises a page keyword dictionary, determining how many page names or keywords in the URL in the streaming data hit the corresponding page keywords in the dictionary when the streaming data is subjected to screening operation; if the dictionary comprises a page dictionary and a page keyword dictionary, screening can be performed based on preset conditions when the streaming data is screened; the preset condition comprises at least one of the following conditions: the page name or URL in the streaming data hits a corresponding page name or URL in a page dictionary, and the page name or URL in the streaming data hits a corresponding page keyword in a page keyword dictionary, that is, the streaming data may be screened by determining that any one or all of preset conditions are met.

In this step, because the query rate per second that the algorithm model Http online service interface needs to accept is very large, the application screens streaming data through a logic rule base, which is composed of rules, and the calculation of the rules is realized by the following two types of dictionaries: the rule is used for indicating whether the dictionary is used as a blacklist or a white list, the blacklist indicates that mutual information of pages and/or keywords in the dictionary is small, and the white list indicates that the mutual information of the pages and/or the keywords in the dictionary is large.

Optionally, the keywords may also be filtered, because too many keywords may increase the task load of streaming data filtering, and may also affect the use efficiency of the model, and therefore, experiments prove that it is a relatively ideal value to reduce the keyword range to about 1000.

Optionally, if the interaction behavior sequence includes a button event of the user on the page, correspondingly, a corresponding button event dictionary exists for the button event, and if the touch operation in the page browsed by the user is determined to be matched with the button event in the dictionary, the screening operation on the streaming data is completed, similarly, if the interaction behavior sequence also includes other content, such as an interaction commodity sequence, and a corresponding dictionary also exists, the streaming data may be screened by using the dictionary.

Therefore, after the stream type data is screened by the rules in the dictionary and the rule logic library, the important data can be written into the message queue, the task load borne by the message queue is reduced, and the back pressure of the stream computing engine or the loss of the important data is reduced.

Optionally, an embodiment of the present application further provides a specific scheme for determining a dictionary based on mutual information, and fig. 8 is a flowchart for determining a dictionary based on mutual information provided in the embodiment of the present application; the process comprises the following steps:

s801, obtaining historical conversation data of a plurality of historical users and intelligent customer service, and extracting corresponding problems according to the historical conversation data to obtain a problem set.

In this step, the problem set is that the historical users communicate with the intelligent customer service in a conversation mode, and then the problems encountered by the users are extracted in the background, at least one problem can be extracted from each communication of the historical users with the intelligent customer service, and the problem set is used for expressing various problems which the users may encounter.

S802, determining a page set and a page keyword set browsed by the plurality of historical users according to the historical browsing page sequence of the plurality of historical users before the historical conversation data.

In the step, historical browse page sequences before historical conversation data corresponding to a plurality of historical users are decomposed, and page sets browsed by the historical users and page keyword sets are further determined; the page set comprises at least one page of a historical user and at least one page keyword.

And S803, for any problem in the problem set, calculating mutual information between each page in the page set and the problem, and mutual information between each page keyword in the page keyword set and the problem, and selecting a target page and a target page keyword corresponding to the problem from the page set and the page keyword set by the calculated mutual information.

In the embodiment of the application, the determined target page can be used for predicting keywords of the user problem, the target page is a page including useful information, that is, all pages corresponding to the conventional browsing behavior of the user are not target pages, for example, pages frequently visited by the user, such as "home page", "my" and the like, and the pages cannot be regarded as target pages.

Specifically, after the mutual information between each page and a problem in the page set is calculated by using the formula (1), the mutual information is sequenced according to the size sequence, further, the mutual information between each page keyword and the problem in the page keyword set is calculated by using the formula (1), the calculated mutual information is also sequenced, and then the page and the page keyword of which the sequencing is positioned at the top N are selected, so that the target page and the target page keyword corresponding to the problem are determined.

S804, fusing target pages corresponding to the problems in the problem set to obtain a page dictionary, and fusing target page keywords corresponding to the problems to obtain a page keyword dictionary.

For example, it is assumed that 100 corresponding questions are extracted from historical session data, 1000 pages are determined from a historical browsing page sequence, and 1 ten thousand page keywords are determined, further, the contribution of each page to each extracted question and the contribution of each page keyword to each extracted question are calculated based on mutual information, after the calculation is completed, for each question, the first 1000 keywords corresponding to the question are found, the first 100 pages, that is, the target page corresponding to the question is the first 100 pages, and the target page keywords are the first 1000 keywords, further, the target pages corresponding to all the questions are fused to obtain a page dictionary, and the target page keywords corresponding to all the questions are fused to obtain a page keyword dictionary.

It should be noted that, in the embodiment of the present application, the number of the target pages and the target page keywords corresponding to each question is not specifically limited, and the number of the questions corresponding to extraction is also not specifically limited, which is merely an example.

Therefore, the method and the device can extract the problems encountered by the user according to the actual historical session data of the user, and determine the contribution of each page and the keywords to the problems encountered by the user by combining the page sequence browsed before the user consults the problems, so that the page dictionary and the page keyword dictionary in the dictionary are accurately determined based on mutual information, the accuracy of determining the dictionary is improved, and the interference of useless data is reduced.

With reference to the foregoing embodiment, fig. 9 is an online computing flowchart of a problem prediction method provided in this embodiment of the present application, and as shown in fig. 9, after a flow computing framework stores a task (streaming data) to a message middleware, the task sequentially passes through a filter, a vectorization module, a tagging module, a sorting module and a process pool to predict a problem encountered by a user, where the filter processes an instant page, such as an active page, based on a filtering logic (a rule logic library), and filters out a keyword or a key page and a rule parameter; vectorization is processed based on a characterization learning model, namely in an embedding expression stage, a serialized page browsing path of a user is coded into a two-dimensional vector, fine tuning training is carried out on the model by utilizing a historical interactive behavior sequence and an extracted problem, and the historical interactive behavior sequence and the extracted problem are labeled, namely classified, through the processing of a basic three-convolution neural network model (problem classification model); the ranking module is used for ranking of questions/knowledge (problems encountered by the user); the process pool is used for buffering/triggering, i.e. triggering, pushed messages to the user.

The tasks obtained by the flow computing framework come from Log services (SLS), the flow computing framework comprises a Service module, a fine operation module and an application configuration center, wherein in the message consumption process of the flow computing framework, the data types output by message consumption can be voices, pictures, characters, videos and the like, further, a corresponding recommendation form is selected based on the output data types to recommend the data to a user, and the recommendation form can be intelligent outgoing calls, messages/short messages, customer Service services and the like; such as the possibility to place the solution in voice form by calling the smart outbound to the user.

Optionally, the corresponding solution may be determined through a preset table according to the problem output by the sorting module, for example, the problem may be converted into card content, as shown in table 1:

TABLE 1

The problem output by the sorting module is inconsistent with the card content finally pushed to the user side, so that a set of card content pushing list needs to be maintained, the mapping relation from the problem encountered by the user to the card ID is included, when the problem is predicted, the corresponding card ID of the problem ID can be searched through the mapping table, and then the card content of the corresponding card ID is called by the background to be pushed to the user side.

Optionally, the method provided by the present application may not be coupled to any push system, and after the output problem is determined, the push system sends a message triggering push, which may be in any other form, such as a card, a file, and the like, and the form may also be added with other designs according to actual requirements, which is not specifically limited in this embodiment of the present application.

According to the method and the device, the relevant indexes can be displayed to the operator, so that the operator can check the problem prediction effect, and further, the operator can make an operation for refining; the related indexes may include a sending user amount, a sending successful user amount, a reach user amount, an effective user amount, an active service session (session) amount, an active service resolution rate, an active service proportion, a reach rate, an effective rate, a click rate, and the like.

Optionally, the sending user amount refers to a pushing user amount received by the pushing system; the sending success user quantity refers to the number of successful pushing solutions, namely, the users receive the pushed contents; the reaching of the user amount means that the user opens the pushed message and sees the pushed card, wherein the message becomes the number that has been read; the effective user amount refers to the number of effective behaviors, such as clicking or replying information, of the user on the basis of seeing the card; the active service session volume refers to the number of session services generated on an intelligent customer service page after a user opens a push message and sees a push card; the proactive service resolution may be a ratio of a successful resolution to a number of possible resolutions, wherein the successful resolution may be a proactive service session amount minus a number of unsuccessful resolutions, and the number of unsuccessful resolutions may include: the online switching amount, the telephone switching amount, the number of unsatisfied sessions, the number of sessions which are recommended for the last time and are not clicked, the number of sessions which have no answer for the last time, the direct connection manual quantity and the like are calculated, and the possible solved number can be the active service session quantity minus the direct connection manual quantity; the active service proportion refers to the ratio of active service session volume to the total service session volume (including active and passive sessions) in the application program; the reach rate refers to the ratio of the amount of reach users to the amount of successful sending users; the effective rate refers to the ratio of the effective user amount to the touch user amount; the click rate refers to the ratio of the amount of users clicked to the amount of users reached.

Optionally, the operation action may include: setting the type and the number of recommendation problems, increasing, deleting, checking and modifying the recommendation problems, controlling the fatigue degree of pushed crowds, editing a pushed template/question case, screening the pushed crowds and the like.

Illustratively, when screening push crowds, messages can be filtered to reach insensitive crowds, negative crowds, or blacklists, that is, for the insensitive crowds and the negative crowds, the messages can be not pushed or pushed less, and user experience is improved.

Optionally, the message pushing opportunity is triggered by a browsing behavior of the user on the application program page, at this time, the application program is in an open state, and the user is reached by the message pushing manner, so that a higher reaching rate and an efficient rate can be ensured.

In another possible implementation manner, the behavior log data may be directly monitored in a manner based on a distributed task system and a handwriting programming logic, and then the problem prediction method shown in fig. 3 is performed, where the method for obtaining the log data is not specifically limited in the embodiment of the present application, and other frames may be used in the embodiment of the present application instead of the stream computation frame, and the method is within the protection scope of the present application as long as the same active problem prediction as the present application can be achieved.

Therefore, the problem prediction method provided by the embodiment of the application has strong pushing timeliness, can be triggered by the browsing behavior of the user, and can output the problem prediction result of the user in second level, so that higher reach rate and efficiency are improved; according to the method and the device, the judgment is carried out by means of a multi-layer and multiple depth prediction algorithm models which are combined with each other without depending on crowd labels, the accuracy of the prediction models is improved, the problem that the input data adopted by the method and the device are dynamic behavior data of users and dynamic order state data and output as intelligent service scenes is solved, the real-time user appeal can be captured more accurately, and the user experience of intelligent service is improved.

Fig. 10 is a schematic flowchart of another problem prediction method provided in an embodiment of the present application, and as shown in fig. 10, a solution may be pushed to a seller according to a problem of a buyer, where the method includes:

s1001, acquiring an interactive behavior sequence of at least one buyer corresponding to the shop within current preset time; wherein the sequence of interaction behaviors comprises a sequence of browsed pages.

In this embodiment, the at least one buyer corresponding to the store may refer to a buyer having a purchase record, a buyer having a comment record, or a buyer having a consultation record in the store.

Optionally, in this step, an interaction behavior sequence of at least one buyer corresponding to the store within the current preset time may be obtained in real time.

S1002, predicting the types of the problems encountered by the buyers based on the problem classification model according to the interaction behavior sequence, and predicting the problems encountered by the buyers based on the problem prediction model corresponding to the types.

In the embodiment of the present application, the problem classification model obtains the type of the problem encountered by each buyer, and the classification of the type may be performed according to the method described in step 302, which is not described herein again, or may be performed according to the number of buyers, that is, the number of passenger flows, for example, when the number of buyers is large, a model with a fast calculation speed may be used, and when the number of buyers is small, a model with a general calculation speed is used to reduce the calculation amount of a preset system.

S1003, determining a solution pushed to a seller corresponding to the shop according to the predicted problem.

In this step, the solutions to be pushed to the sellers corresponding to the stores are determined, and the solutions to be pushed to the sellers are selected based on the types and sizes of the stores, and if the stores are large, the solutions to be pushed to the stores are selected to be preferentially pushed to the buyers having a high priority, such as member clients or buyers having good shopping credit, and if the stores are small, the solutions to be pushed to the stores are selected to be pushed to all the buyers.

It should be noted that the pushing form and the searching manner of the solution are similar to those of step 303 in the above embodiment, and are not described herein again, and refer to the description of step 303. Alternatively, the buyer-specific solution and the seller-specific solution may be different for the same issue, for example, for a seller, if there are refund issues for multiple buyers, the seller may be pushed with a copy: there may be refund issues with multiple buyers, please click on here to view the solution.

Therefore, the browsing page sequence of at least one buyer in the current preset time can be obtained in the embodiment of the application; predicting the types of the problems encountered by the buyers based on the problem classification model, and actively predicting the problems encountered by the buyers based on the problem prediction model corresponding to the types; furthermore, according to the predicted problems, the solutions pushed to the sellers corresponding to the shops are determined, and the efficiency and the accuracy of solving the user problems by the sellers are improved.

Optionally, predicting the type of the problem encountered by each buyer based on the problem classification model according to the interaction behavior sequence, and predicting the problem encountered by each buyer based on the problem prediction model corresponding to the type, including:

For any buyer, inputting the feature vector corresponding to the interaction behavior sequence into the problem classification model, and determining the type of the problem encountered by the buyer;

if the type is an operation and sales class or an account class, processing the interaction behavior sequence of the buyer based on a sequence model to obtain a corresponding problem;

and if the type is a transaction type, processing the interaction behavior sequence of the buyer and the transaction information corresponding to the buyer based on a deep interest evolution network model to obtain a corresponding problem.

Optionally, historical conversation data of the historical seller and the intelligent customer service can be obtained, and corresponding problems are extracted according to the historical conversation data;

according to the historical interactive behavior sequence of the historical seller corresponding to the historical buyer in front of the historical session data and the extracted problems, constructing a training data set corresponding to a sequence model, and training the sequence model; and/or the presence of a gas in the gas,

and constructing a training data set corresponding to the deep interest evolution network model according to the historical interaction behavior sequence, transaction information and extracted problems of the historical buyer of the historical seller before the historical conversation data, and training the deep interest evolution network model.

Optionally, the obtaining of the interaction behavior sequence of the at least one buyer corresponding to the store within the current preset time includes:

generating streaming data containing an interaction behavior sequence according to an interaction log of at least one buyer corresponding to the shop within the current preset time through a stream computing frame;

and screening the streaming data based on a rule logic base, and putting the screened streaming data into a message queue, wherein the streaming data in the message queue is used for processing the problem classification model.

the dictionary comprises a page dictionary and/or a page keyword dictionary;

the dictionary is determined through a history page browsed by a history buyer and a corresponding problem;

the rule is used for indicating whether the dictionary is used as a black list or a white list, and/or the screening operation of the page sequence is completed when how many pages in the page sequence browsed by the buyer are matched with the dictionary.

Optionally, the method further includes:

acquiring historical conversation data of a plurality of historical sellers and intelligent customer service, and extracting corresponding problems according to the historical conversation data to obtain a problem set;

determining a page set browsed by the historical buyers and a page keyword set according to the historical browsing page sequence of the historical buyers before the historical conversation data of the multiple historical sellers;

And fusing target pages corresponding to all the problems in the problem set to obtain a page dictionary, and fusing target page keywords corresponding to all the problems to obtain a page keyword dictionary.

The implementation principle and technical effect of the above embodiment are similar to those of the embodiment of the dependent solution corresponding to fig. 3, and are not described herein again.

In response to the problem prediction method, an embodiment of the present application provides a problem prediction apparatus, and fig. 11 is a schematic structural diagram of the problem prediction apparatus provided in the embodiment of the present application, where the apparatus includes:

a first obtaining module 1101, configured to obtain an interaction behavior sequence of a user within a current preset time; wherein the sequence of interaction behaviors comprises a sequence of browsed pages;

a first prediction module 1102, configured to predict, according to the interaction behavior sequence, a type of a problem encountered by the user based on a problem classification model, and predict, based on a problem prediction model corresponding to the type, a problem encountered by the user;

a first determining module 1103, configured to determine a solution to be pushed to the user according to the predicted problem.

The problem prediction apparatus provided in the embodiment of the present application may be used to implement the technical solutions in the embodiments shown in fig. 1 to 9, and the implementation principles and technical effects thereof are similar, and this embodiment is not described herein again.

For example, an embodiment of the present application further provides a problem prediction apparatus, and fig. 12 is a schematic structural diagram of another problem prediction apparatus provided in the embodiment of the present application, where the apparatus includes:

a second obtaining module 1201, configured to obtain an interaction behavior sequence of at least one buyer corresponding to the store within a current preset time; wherein the sequence of interaction behaviors comprises a sequence of browsed pages;

the second prediction module 1202 is configured to predict the type of the problem encountered by each buyer based on the problem classification model according to the interaction behavior sequence, and predict the problem encountered by each buyer based on the problem prediction model corresponding to the type;

a second determining module 1203, configured to determine, according to the predicted problem, a solution to be pushed to a seller corresponding to the store.

The problem prediction apparatus provided in the embodiment of the present application may be used to implement the technical solutions in the embodiments shown in fig. 1 to 10, and the implementation principles and technical effects thereof are similar, and this embodiment is not described herein again.

Fig. 13 is a schematic structural diagram of an electronic device according to an embodiment of the present application. As shown in fig. 13, the electronic device of the present embodiment may include:

at least one processor 1301; and

A memory 1302 communicatively coupled to the at least one processor;

wherein the memory 1302 stores instructions executable by the at least one processor 1301, the instructions being executable by the at least one processor 1301 to cause the electronic device to perform the method according to any of the embodiments described above.

Alternatively, the memory 1302 may be separate or integrated with the processor 1301.

For the implementation principle and the technical effect of the electronic device provided by this embodiment, reference may be made to the foregoing embodiments, which are not described herein again.

The embodiment of the present application further provides a computer-readable storage medium, in which computer-executable instructions are stored, and when a processor executes the computer-executable instructions, the method described in any one of the foregoing embodiments is implemented.

The present application further provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, the computer program implements the method described in any of the foregoing embodiments.

In the technical scheme of the application, the collection, storage, use, processing, transmission, provision, disclosure and other processing of the related user data and other information all accord with the regulations of related laws and regulations and do not violate the good customs of the public order.

In the several embodiments provided in the present application, it should be understood that the disclosed apparatus and method may be implemented in other ways. For example, the above-described apparatus embodiments are merely illustrative, and for example, the division of the modules is only one logical functional division, and other divisions may be realized in practice, for example, a plurality of modules may be combined or integrated into another system, or some features may be omitted, or not executed.

The integrated module implemented in the form of a software functional module may be stored in a computer-readable storage medium. The software functional module is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device) or a processor to execute some steps of the methods described in the embodiments of the present application.

It should be understood that the Processor may be a Central Processing Unit (CPU), other general purpose Processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), etc. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of the methods disclosed in the incorporated application may be directly implemented by a hardware processor, or may be implemented by a combination of hardware and software modules in the processor. The memory may comprise a high-speed RAM memory, and may further comprise a non-volatile storage NVM, such as at least one disk memory, and may also be a usb disk, a removable hard disk, a read-only memory, a magnetic or optical disk, etc.

The storage medium may be implemented by any type or combination of volatile or non-volatile memory devices, such as Static Random Access Memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic or optical disks. A storage media may be any available media that can be accessed by a general purpose or special purpose computer.

An exemplary storage medium is coupled to the processor such the processor can read information from, and write information to, the storage medium. Of course, the storage medium may also be integral to the processor. The processor and the storage medium may reside in an Application Specific Integrated Circuits (ASIC). Of course, the processor and the storage medium may reside as discrete components in an electronic device or host device.

It should be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising a … …" does not exclude the presence of another identical element in a process, method, article, or apparatus that comprises the element.

The above-mentioned serial numbers of the embodiments of the present application are merely for description and do not represent the merits of the embodiments.

Through the above description of the embodiments, those skilled in the art will clearly understand that the method of the above embodiments can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware, but in many cases, the former is a better implementation manner. Based on such understanding, the technical solutions of the present application may be embodied in the form of a software product, which is stored in a storage medium (such as ROM/RAM, magnetic disk, optical disk) and includes instructions for enabling a terminal device (such as a mobile phone, a computer, a server, an air conditioner, or a network device) to execute the method according to the embodiments of the present application.

The above description is only a preferred embodiment of the present application, and not intended to limit the scope of the present application, and all the equivalent structures or equivalent processes that can be directly or indirectly applied to other related technical fields by using the contents of the specification and the drawings of the present application are also included in the scope of the present application.

Claims

1. A problem prediction method, characterized in that the method comprises:

predicting the type of the problem encountered by the user based on a problem classification model according to the interaction behavior sequence, and predicting the problem encountered by the user based on a problem prediction model corresponding to the type;

determining a solution to push to the user based on the predicted problem.

2. The method of claim 1, wherein predicting, according to the sequence of interaction behaviors, a type of problem encountered by the user based on a problem classification model and predicting a problem encountered by the user based on a problem prediction model corresponding to the type comprises:

3. The method of claim 2, further comprising:

4. The method of claim 3, wherein the sequence model is a natural language model for processing sequences; training the sequence model, including:

5. The method according to any one of claims 1 to 4, wherein acquiring the sequence of the interaction behaviors of the user within the current preset time comprises:

6. The method of claim 5, wherein screening the streaming data based on a rule logic base and placing the screened streaming data in a message queue comprises:

the dictionary comprises a page dictionary and/or a page keyword dictionary;

the dictionary is determined through historical pages browsed by historical users and corresponding problems.

7. The method of claim 6, further comprising:

8. A problem prediction method, comprising:

9. An electronic device, comprising:

at least one processor; and

a memory communicatively coupled to the at least one processor;

wherein the memory stores instructions executable by the at least one processor to cause the electronic device to perform the method of any one of claims 1-8.

10. A computer-readable storage medium having stored thereon computer-executable instructions that, when executed by a processor, perform the method of any one of claims 1-8.

11. A computer program product comprising a computer program, characterized in that the computer program realizes the method according to any of claims 1-8 when executed by a processor.