CN115147130A - Problem prediction method, device, storage medium and program product - Google Patents

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, device, storage medium and program product

technical field

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

Background technique

With the continuous development of artificial intelligence technology, intelligent customer service robots have the ability to ask questions and dialogues in the knowledge base, and can answer user questions smoothly and accurately.

However, the above-mentioned customer service robots usually accept user consultation passively, requiring users to find the entrance of the customer service robot and express service demands, which is time-consuming and labor-intensive, wastes resources, and it is difficult to find user problems in a timely and effective manner, and solve user problems in the first time. The user experience is relatively poor. Difference.

SUMMARY OF THE INVENTION

The main purpose of the embodiments of the present application is to provide a problem prediction method, device, storage medium and program product, so as to perform real-time problem and intention prediction and improve the service experience of users.

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

Acquire the interactive behavior sequence of the user within the current preset time; wherein, the interactive behavior sequence includes the browsed page sequence;

According to the interactive behavior sequence, the type of the problem encountered by the user is predicted based on the problem classification model, and the problem encountered by the user is predicted based on the problem prediction model corresponding to the type;

Based on the predicted problem, a solution to push to the user is determined.

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

Input the feature vector corresponding to the interactive behavior sequence into the problem classification model to determine the type of problem encountered by the user;

If the type is a marketing type or an account type, the user's interactive behavior sequence is processed based on the sequence model to obtain the corresponding question;

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

Optionally, also include:

Obtain historical conversation data between historical users and intelligent customer service, and extract corresponding questions according to the historical conversation data;

According to the historical interaction behavior sequence of the historical user before the historical session data and the extracted questions, a training data set corresponding to the sequence model is constructed, and the sequence model is trained; and/or,

According to the historical interaction behavior sequence, transaction information and extracted questions of the historical user before the historical session data, a training data set corresponding to the deep interest evolution network model is constructed, and the deep interest evolution network model is trained.

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

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

Inputting the word segmentation sequence into the natural language model, and pre-training the natural language model with predicting the next sentence as a training target;

Taking the extracted questions as supervision signals, fine-tuning training is performed on the pre-trained natural language model according to the historical interaction behavior sequence.

Optionally, obtain the interactive behavior sequence of the user within the current preset time, including:

Through the stream computing framework, according to the user's interaction log within the current preset time, the stream data containing the interaction behavior sequence is generated;

The streaming data is filtered based on the rule logic library, and the filtered streaming data is put into a message queue, and the streaming data in the message queue is used for processing by the problem classification model.

Optionally, filter the streaming data based on the rule logic library, and put the filtered streaming data into a message queue, including:

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

Wherein, the dictionary includes a page dictionary and/or a page keyword dictionary;

The dictionary is determined by historical pages browsed by historical users and corresponding questions;

The rule is used to indicate whether the dictionary is used as a blacklist or a whitelist, and/or the number of pages in the page sequence browsed by the user that matches the dictionary is used to complete the screening operation on the page sequence.

Optionally, also include:

Obtaining historical conversation data of multiple historical users and intelligent customer service, extracting corresponding questions according to the historical conversation data, and obtaining a question set;

determining a page set and a page keyword set browsed by the plurality of historical users according to the sequence of historical pages viewed by the plurality of historical users before the historical session data;

For any question in the question set, calculate the mutual information between each page in the page set and the question, and the mutual information between each page keyword in the page keyword set and the question, and calculate the mutual information obtained. The information selects the target page and target page keyword corresponding to the question from the page set and the page keyword set;

The target pages corresponding to each question in the question set are fused to obtain a page dictionary, and the target page keywords corresponding to each question are fused to obtain a page keyword dictionary.

In a second aspect, the embodiments of the present application also provide a problem prediction method, including:

Obtain the interactive behavior sequence of at least one buyer corresponding to the store within the current preset time; wherein, the interactive behavior sequence includes the browsed page sequence;

According to the interactive behavior sequence, the type of the problem encountered by each buyer is predicted based on the problem classification model, and the problem encountered by each buyer is predicted based on the problem prediction model corresponding to the type;

According to the predicted problem, a solution to be pushed to the seller corresponding to the store is determined.

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, and the instructions are executed by the at least one processor to cause the electronic device to perform the method according to any one of the above aspects.

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

In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, which implements the method described in any one of the foregoing aspects when the computer program is executed by a processor.

The problem prediction method, device, storage medium, and program product provided by the embodiments of the present application can obtain the interactive behavior sequence of the user within the current preset time; wherein, the interactive behavior sequence includes the browsed page sequence; further, according to the interactive behavior Sequence, predicts the types of problems encountered by users based on the problem classification model, and predicts the problems encountered by users based on the problem prediction model corresponding to the types; further, according to the predicted problems, determine the solution to be pushed to the user. This application can classify the page sequence corresponding to the user's browsing application within a preset time, determine the type corresponding to the page sequence, and then select a model matching the type to predict the problem, and accurately predict the user's current possible existing problems. Problems, timely detection of user problems, improve the accuracy, real-time and effectiveness of problem prediction, and can also recommend solutions to users, solve user problems in the first time, and improve user service experience.

Description of 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 diagram of an application scenario involved in an embodiment of the present application;

FIG. 2 is a schematic diagram showing an interface of a terminal device according to an embodiment of the present application;

3 is a schematic flowchart of a problem prediction method provided by an embodiment of the present application;

FIG. 4 is a schematic structural diagram of a stream computing framework provided by an embodiment of the present application;

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

6 is a schematic diagram of training a deep interest evolution network model provided by an embodiment of the present application;

7 is a schematic diagram of pre-training a natural language model provided by an embodiment of the present application;

8 is a flowchart of determining a dictionary based on mutual information provided by an embodiment of the present application;

FIG. 9 is an online calculation flowchart of a problem prediction method provided by an embodiment of the present application;

10 is a schematic flowchart of another problem prediction method provided by an embodiment of the present application;

FIG. 11 is a schematic structural diagram of a problem prediction apparatus provided by an embodiment of the present application;

FIG. 12 is a schematic structural diagram of another problem prediction apparatus provided by 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.

Specific embodiments of the present application have been shown by the above-mentioned drawings, and will be described in more detail hereinafter. These drawings and written descriptions are not intended to limit the scope of the concepts of the present application in any way, but to illustrate the concepts of the present application to those skilled in the art by referring to specific embodiments.

Detailed ways

Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. Where the following description refers to the drawings, the same numerals in different drawings refer to the same or similar elements unless otherwise indicated. The implementations described in the illustrative examples below are not intended to represent all implementations consistent with this application.

First, the terms involved in this application are explained:

Stream computing: refers to a computing architecture that can acquire massive data from different data sources in real time, and obtain valuable information through real-time analysis and processing, which is used for real-time analysis of large-scale flowing data in the process of changing motion. analysis to capture potentially useful information and send the results to the next compute node.

Stream computing engine: refers to a real-time computing engine developed based on an open-source distributed computing framework to implement stream computing.

Recommendation system: refers to the use of e-commerce websites to provide customers with product information and suggestions, to help users decide what products they should buy, and to simulate salespeople to help customers complete the purchase process, such as providing users with personalized sorting based on users’ historical preferences and constraints Item recommendation list, that is, recommendation results can be generated according to user preferences, commodity characteristics, user-commodity transactions and other environmental factors (such as time, season, location, etc.). The recommended items can include movies, books, restaurants, news Items, etc. In this application, the personalized items may be questions in the customer service system, and may also be referred to as intents or scenarios.

C-end users: In the e-commerce scenario, it usually refers to consumers or buyers.

Through the development in recent years, the intelligent customer service robot has the knowledge base Q&A and dialogue capabilities, and can answer user questions smoothly and accurately. Association conjecture.

However, the above-mentioned customer service robots usually accept user consultation passively, requiring users to find the entrance of the customer service robot and express service demands, which is time-consuming and labor-intensive, wastes resources, and it is difficult to find user problems in a timely and effective manner, and solve user problems in the first time. The user experience is relatively poor. Difference.

However, in the process of using the application, various operations will leave behavioral traces. Therefore, in some technologies, the user's possible feelings can be extracted based on the user portrait or the detected user log generated when the user uses the application. Interested content actively recommends different messages to different users. For example, a push system based on crowd tags can actively push different messages, publicity and promotion texts, etc. to different groups of people.

Exemplarily, the push system can run structured query language (Structured Query Language, SQL) in the background data table to circle people that meet the set conditions, or a tag system based on demographics and interest models, through a combination of crowd tags. To manually filter target users, and then push the message.

However, the above-mentioned crowd tag-based push system actively pushes messages by relying on crowd tags and attributes to mark users. The development of such crowd tagging systems requires a long period of time, and crowd tags supported by third parties often do not On the other hand, the tag system based on demographics and interest models is a relatively long-term attribute tag and will not change significantly in the short term. Therefore, the accurate information of messages pushed to users by tagging performance and real-time performance are poor.

It should be noted that the above push systems are mostly one-time push for the purpose of event publicity and promotion, and are not instant push scenarios customized for intelligent customer service robots. Poor, and the timing of message push is not the first time the problem occurs. Therefore, after the target is pushed through the above method, there is often a certain lag, which greatly reduces the user's willingness to open the push message.

However, since the robot has the ability to predict problems, in the intelligent service, it can obtain user logs in real time, actively predict the user's problems, and send a push message to the user at the first time, reminding the user to click the push message to help solve the problem. .

Therefore, for intelligent customer service robots, the ability of stream computing can be used to make real-time predictions on problems that users may encounter, and push relevant solutions, thereby improving the quality and temperature of services. For example, when a C-end user is using an electronic product, the system will record the Uniform ResoureLocator (URL) of the user's visit to the browsing page. The URL of each page corresponds to an operation of the user. Sorting can constitute the user's behavioral trajectory, which includes the user's immediate intentions and needs. Further, based on the user's behavioral trajectory, as well as the monitoring and prediction of abnormal behaviors, the problems that the user may encounter can be solved. Make real-time predictions to recommend relevant solutions.

In view of this, the embodiment of the present application provides a problem prediction method, which can process the log data that the user browses on the page in the application to process the message trigger source of stream computing, and at the same time use the algorithm model to analyze the possibility of the user in the application. The type of the problem is encountered, and then a corresponding model is selected based on the type of the problem to predict the problem and intent in real time. Exemplarily, FIG. 1 is a schematic diagram of an application scenario involved in an embodiment of the present application. As shown in FIG. 1 , the problem prediction method provided by the embodiment of the present application can be applied to the application scenario shown in FIG. 1 , and the application scenario includes: a terminal device 101 , an analysis platform 102 , and a processing platform 103 ; specifically, the analysis platform 102 The page sequence of an application program browsed by the terminal device 101 within 30 minutes can be obtained, and further, based on the obtained page sequence corresponding to the terminal device 101 browsed within 30 minutes, the problem classification model is used to determine the type of the page sequence , and 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 the problem 1 that may be encountered, and send the problem 1 to the processing platform 103. Further, the processing platform 103 According to the problem 1, the solution 1 to be pushed to the user is determined, and the determined solution 1 is sent to the terminal device 101 for display for the user to view.

It can be understood that, at preset time intervals, the processing platform 103 may adjust the solution corresponding to question 1 based on the user's feedback information, so that the solution is more in line with the user's needs and improves the user experience.

Among them, the analysis platform 102 includes a large amount of data, which can be classified, and then determine the problem prediction model corresponding to each type and save it for problem prediction, so that it can be applied to a variety of scenarios, using the analysis platform 102 as the electrical Take the e-commerce platform as an example, it can be divided according to the type of e-commerce platform. Each type corresponds to the problem prediction model. If there are 3 types, it corresponds to the problem prediction model 1, the problem prediction model 2 and the problem prediction model 3. Each The prediction model can predict one or more problems of a corresponding type. The embodiments of the present application do not specifically limit the categories of classification and the types and quantities of the corresponding problem prediction models, and the above are only examples.

It should be noted that the above steps of using the problem classification model to determine the type of page sequence, calling the problem prediction model corresponding to the type to predict the problems encountered by the user, and determining the solution to be pushed to the user according to the problem are just examples. It should be noted that the processing may also be performed only through one platform such as the cloud, or more platforms and more modules may be split for processing, which is not specifically limited in this embodiment of the present application.

Optionally, the log corresponding to a page of an application program browsed by the terminal device 101 within 30 minutes may be obtained in the following manner. FIG. 2 is a schematic diagram of an interface display of a terminal device provided by an embodiment of the present application, as shown in FIG. 2 . , the terminal device can obtain browsing log data according to the application program interface opened by the user and the touch operation on the application program interface. Go to the personal user interface, click the "Refund/After-Sales" button again through the touch operation in My Order, browse the data related to the refund/after-sales, and then, the terminal device obtains the browsing log data.

Further, after acquiring the browsing log data, the terminal device can send the data to a log storage module, and the log storage module can be any device or platform capable of storing log data, such as a log service data center or a data warehouse. Based on the solution provided in this application, stream computing and other methods can be used to convert log data into streaming data in real time to initiate downstream active service processes. Among them, the log service data center or data warehouse persistent storage, when triggered by the flow calculation of this solution, can achieve a second-level delay.

Optionally, the active service process can determine the problems that users may encounter, such as "users need to understand the refund or after-sales process", and then push corresponding solutions such as "refund process/after-sales process" to users, Help users solve possible problems in a timely manner.

Therefore, the problem prediction method provided in the embodiment of the present application can proactively predict the user's current possible problems without the need for the user to ask questions to the customer service robot, save time, and at the same time discover the user's problems in a timely manner. When encountering problems, the algorithm model is used to determine the types of problems that users may encounter, and then the corresponding models are selected based on the types to predict the problems, which improves the accuracy and real-time performance of problem prediction, and can also actively push the corresponding solutions to Users, realize companion service and improve user service experience.

The technical solutions of the present application will be described in detail below with specific examples. The following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The 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 by an embodiment of the present application. This embodiment can be applied to any device capable of implementing data processing, such as a server and the like. As shown in Figure 3, the method may include:

S301. Acquire a sequence of interactive behaviors of a user within a current preset time; wherein, the sequence of interactive behaviors includes a sequence of browsed pages.

In this embodiment of the present application, the interaction behavior sequence is used to reflect the interaction behavior between the user and the application program in the terminal device, including a browsed page sequence, and the page sequence includes information of multiple pages browsed by the user within the current preset time. The information of the page includes a page name and/or a page URL; the page name may include, for example, an order details page, a help page, a product details page, and the like.

Optionally, in this step, the interactive behavior sequence of the user within the current preset time may be acquired in real time. The preset time refers to the time set by the preset system for real-time processing of the collected data, and the preset system can be a system set in the cloud for data processing, so that the user's browsing log data can be Get instant processing, and then improve the accuracy of the forecast, such as the preset time can be 30 minutes.

Wherein, the interactive behavior sequence may be a sequence of pages obtained by the user within the current preset time and meeting preset requirements, and the preset requirements may be a preset number, a specified page, or other requirements , which is not specifically limited in the embodiments of the present application.

Exemplarily, in the application scenario of FIG. 1 , the page sequences corresponding to 30 real-time shopping pages browsed by the terminal device 101 in a certain application program in the current 30 minutes may be obtained, that is, 30 URL records.

S302. Predict the type of the problem encountered by the user based on the problem classification model according to the interactive behavior sequence, and predict the problem encountered by the user based on the problem prediction model corresponding to the type.

In this embodiment of the present application, the types of questions can be classified according to the nature of the input data. For example, when the input data is only an interactive behavior sequence, a model that only processes the sequence can be used. For example, the input data is an interactive behavior sequence and Other information, such as order status, requires a model that can handle both sequences and other information.

Specifically, the types include at least the following two types: predicting problems based only on interaction behavior sequences, and predicting problems based on interaction behavior sequences and other information; and calling a corresponding model to predict problems encountered by users according to the types. The problems may include rights protection problems, product recommendations, logistics problems, store evaluation information, etc., which are not specifically limited in this embodiment of the present application.

Exemplarily, the present application can only call the corresponding model based on the interactive behavior sequence to predict the problems encountered by the user. The problem encountered by the user can also be called based on the interactive behavior sequence and other information to call the corresponding model to predict the problem encountered by the user; the other information may include order status, logistics status, and optional problem scenarios, etc., which are not addressed in this embodiment of the application The specific limitation is that the more types and contents other information contains, the higher the accuracy of the prediction problem. The selected problem scenario is used as the input of the model, and further, the corresponding model is called to obtain the probability that the user selects each problem, that is, the problem encountered by the user is predicted.

Optionally, it can also be classified according to the nature of the output data. For example, the output data that has little influence on the user, such as the data related to the user's preferences, can use a model with general accuracy but fast speed, and the output data that has a great influence on the user, such as For transaction data, use a model with higher accuracy, or classify according to the amount of calculation of the output data. A model with average accuracy can reduce the burden on the preset system.

It should be noted that the specific models corresponding to the problem classification model and the problem prediction model are not limited in the embodiments of the present application, and the problem classification model and the problem prediction model can satisfy the following conditions: when using the problem classification model to classify the interactive behavior sequence After that, the problem prediction model corresponding to the type can be called to predict the problems that users may encounter.

S303. Determine a solution to be pushed to the user according to the predicted problem.

Specifically, according to the predicted problem, the user's solution can be obtained by matching the corresponding message template, and the solution can be pushed to the user. In this embodiment of the present application, the solution can also be obtained in other ways. The solution of the user is described above, which is not specifically limited in this embodiment of the present application.

In this step, the solution may refer to the solution that is proactively pushed to the user for the predicted problem that the user may encounter. The solution may be stored in the look-up table in advance and called directly during use. Each problem has a corresponding solution, and the corresponding solution can be found in the lookup table through the id corresponding to the problem.

Optionally, the solution can be a copy, link, card, etc., which can be pushed to the user through messages, text messages, phone calls, etc. in the application, for example, a link for refund can be recommended to the user in the form of a message.

Exemplarily, taking a travel platform as an example, through the user browsing the page information of the application program corresponding to the travel platform, if the problem encountered by the user is predicted to be a rights protection problem based on the corresponding problem prediction model, it will actively report to the user's terminal device. The display interface recommends the process of rights protection; if the user’s problem is predicted to be a travel route selection problem based on the corresponding problem prediction model, it will actively recommend the current season’s travel route, travel agency and travel strategy to the display interface of the user’s terminal device etc.; if the problem encountered by the user is predicted to be a ticketing problem based on the corresponding problem prediction model, it will proactively recommend a ticket purchase process, a ticket change process or a ticket refund process to the display interface of the user's terminal device.

Take a takeaway platform as an example, through the user browsing the webpage information of the application program corresponding to the takeaway platform, if the problem encountered by the user is predicted to be a meal ordering problem based on the corresponding problem prediction model, the user will actively push the display interface of the user's terminal device. Possible selection of takeaway-related information; taking an e-commerce platform as an example, by counting the webpage information and real-time product reviews, service data, etc. If the problem is a shopping problem, it will actively push the relevant information of the store and merchant that may be selected to the display interface of the user's terminal device.

Therefore, the embodiment of the present application can obtain the page sequence browsed by the user in the current preset time; predict the type of problems encountered by the user based on the problem classification model, and actively predict the problem encountered by the user based on the problem prediction model corresponding to the type. problems; further, according to the predicted problems, determine a solution to push to the user. This application can classify the page sequence corresponding to the user's browsing application within a preset time, determine the type corresponding to the page sequence, and then select a model matching the type to predict the problem, and accurately predict the user's current possible existing problems. Problems, timely detection of user problems, improve the accuracy, real-time and effectiveness of problem prediction, and can also recommend solutions to users, solve user problems in the first time, and improve user service experience.

Based on the above considerations, the problem prediction method provided by the present application may be implemented based on a flow computing framework. FIG. 4 is a schematic structural diagram of a flow computing framework provided by an embodiment of the present application. As shown in FIG. 4 , the flow computing framework is A hierarchical user intent prediction framework can predict the problems that users may encounter based on the user's browsing log data, and improve the accuracy of problem prediction. Specifically, the flow computing framework includes two components, namely a flow computing message trigger source consisting of a flow computing custom function module, an algorithm model Http online service module (referred to as algorithm service) and a rule logic library, and a recall module consisting of a flow computing message trigger source. A message consumption prediction component composed of (Matching), ranking module (Ranking), fatigue control, push control and other modules; the stream computing message trigger source integrates stream computing engine and message middleware (MetaQ) and other middleware, message consumption Prediction is used to call various algorithm models hierarchically, such as calling the recall algorithm model library and the algorithm models in the sorting algorithm model library to predict the problems that users may encounter.

Exemplarily, the user log of the user's browsing page is processed into stream data of key-value pairs by the stream computing engine. Further, the stream computing user-defined function module calls the algorithm in the Http online service module of the algorithm model. The obtained key pages/keywords, as well as the rule logic library, can filter a large amount of streaming data, and the filtered streaming data is put into the message middleware as a message for message consumption prediction.

Further, in the process of message consumption prediction, the message passes through the recall module and the sorting module in turn to predict the problems that the user may encounter, and then based on the push control module, the problems are fed back to the offline data computing (scheduling) service, and the feedback The content also includes a log table and a push table, the log table is used to store user logs, the push table is used to update mutual information, and the mutual information is used to measure "the appearance of a certain keyword/key page" for "knowledge". The amount of information contributed by the "appearance of the point", and the "knowledge point" is the problem encountered by the user, thus forming a closed loop.

Among them, the purpose of current limit is to reduce the accumulation of messages. When the amount of messages is greater than the threshold, some messages can be discarded. The fatigue control module is used to set the number of pushes to the user, so as to reduce the user's disgust caused by too many pushes. mood.

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

Input the feature vector corresponding to the interactive behavior sequence into the problem classification model to determine the type of problem encountered by the user;

If the type is a marketing type or an account type, the user's interactive behavior sequence is processed based on the sequence model to obtain the corresponding question;

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

In this embodiment of the present application, the feature vector may refer to a two-dimensional vector obtained after vectorization, and may also be referred to as an embedded representation vector. Optionally, a word2vec model can be used to vectorize the sequence of interactive actions. Exemplarily, the user's browsing log can be obtained in real time by accessing a certain application, and the user's serialized page browsing path (ie, interactive behavior sequence), such as page 1 to page N, can be determined. In the embedding representation stage, the user's serialized page browsing path is encoded into a two-dimensional vector, and then the problem is classified according to the two-dimensional vector.

In this step, the problem classification model is used to classify the problems encountered by the user, and predict whether the current user has one or more of the three types of problems: account, transaction, and marketing. By labeling the problems, That is, the user's problem is labeled to call the corresponding problem prediction model to predict the user's problem based on the interactive behavior sequence; the problem prediction model includes: language model-based problem prediction models such as sequence models, based on deep learning click rate ( The problem prediction model of the Click Through Rate (CTR) model is such as the Deep Interest Evolution Network (DIEN) model, the sequence model is used to solve the problem prediction under the marketing and account labels; the DIEN model is used to solve the transaction The problem prediction under the label is to do a secondary ranking of the problems encountered by users.

Among them, the introduction of the CTR model in the embodiment of the present application is to convert the problem prediction problem into a matching problem of judging "user (User) and problem (item)", that is, judging whether the current user will click (1) or will not click (0). ), and then solve it by scoring 0/1. Specifically, the feature vector input by the CTR model includes user page browsing behavior sequence, discrete features and continuous features; the discrete features such as order, logistics, service status, problem id etc., the continuous features such as the order price, the click rate of the item, etc., input the feature vector into the CTR model, the output "predicted click rate" of the model can be obtained, which is a score of 0 to 1, that is, whether the user has the item or not. When a click behavior is generated, the click is 1, and the non-click is 0. Further, the question with the corresponding high predicted score in the question can be selected as the question to be pushed to the user first.

In view of the different data forms of different question labels, different question classification models are used for prediction. If the type is marketing or account type, the problems encountered by users are predicted based on the sequence model. The sequence model can be The natural language model may also be any other model capable of processing sequences, which is not specifically limited in the embodiments of this application. For example, the sequence model is a Bert-MM-1 model, and if the type is a transaction class, it is based on deep interest Evolutionary network models make predictions about problems encountered by users.

Among them, the type of problem predicted based only on the sequence of interactive behavior is marketing or account type, and the type of problem predicted based on the sequence of interactive behavior and other information is transaction type; Red envelopes, coupons, etc.; the account category refers to the problems encountered by users in modifying personal information, including modifying the delivery address, modifying the mobile phone number bound to the account, etc.; the transaction category refers to the problems encountered by users in transactions, including How to apply for a refund, query the logistics status, query the order status, etc., the embodiment of this application does not specifically limit the number of problems included in each type.

Specifically, the problem classification model vectorizes the interactive behavior sequence, that is, converts the user's interactive behavior sequence into a two-dimensional vector encoded by the page browsing path, and inputs the two-dimensional vector into the corresponding convolutional neural network. In the network (Convolutional Neural Network, CNN), the serialized implicit pattern is extracted through CNN, and the classification corresponding to the problem is output as the prediction result; further, based on the type corresponding to the problem, the corresponding problem prediction model is selected, and interactive The action sequence is used as the input to the model, and the output of the probability that the user selects each question is obtained. For example, FIG. 5 is a schematic diagram of a process for determining problem classification provided by an embodiment of the application. As shown in FIG. 5 , the implicit pattern of the serialized embedded representation vector can be extracted through CNN, that is, after convolution, maximum pooling, The fully connected neural network and the classifier are processed to obtain the classification of the problem as the prediction result.

Optionally, the process of vectorizing the interactive behavior sequence can be unsupervised. Specifically, during the coding model training process, the “semantic information” can be mined unsupervised based on the page browsing paths of a large number of users and saved as the coding model. , the encoding model can be understood as the word2vec model, that is, the page browsing path and the corresponding two-dimensional vector are input into the word2vec model for training, and the trained word2vec model is obtained.

Therefore, since the data forms corresponding to the interactive behavior sequences are different, this embodiment of the present application can call a corresponding problem prediction model based on the type of problems encountered by the user and use the interactive behavior sequence to predict the problem. There are corresponding data types for different data types. The processing method enables different data to predict user problems and improves the flexibility of problem prediction scenarios.

Optionally, historical conversation data between historical users and intelligent customer service can also be obtained, and corresponding questions can be extracted according to the historical conversation data;

According to the historical interaction behavior sequence of the historical user before the historical session data and the extracted questions, a training data set corresponding to the sequence model is constructed, and the sequence model is trained; and/or,

According to the historical interaction behavior sequence, transaction information and extracted questions of the historical user before the historical session data, a training data set corresponding to the deep interest evolution network model is constructed, and the deep interest evolution network model is trained.

In the embodiment of this application, historical users may refer to users who have browsing records in the application within a certain period of time and who have actively consulted with questions, such as user 1 who has browsed the refund page of an application and consulted about refunds, Historical session data can refer to the session data of historical users and intelligent customer service for question exchanges, such as a series of conversation contents in which user 1 and customer service consulted about the refund process of an item. According to historical session data, based on the natural language processing model, Corresponding questions can be extracted.

Specifically, the embodiment of the present application constructs a corresponding training data set based on the historical interaction behavior sequence of historical users before historical session data and the extracted questions. The extracted questions are the real results corresponding to the training data, that is, labels, which are used to improve training. accuracy.

Exemplarily, the sequence model may be any model capable of processing sequence data, such as a natural language model. The training set for building a sequence model can be obtained by the following methods: obtaining the interactive behavior sequence of at least one user in a certain application program interface within a period of time as feature data, and then based on the conversation data between each user and intelligent customer service during this period, The corresponding problem is extracted as the label corresponding to the feature data. The feature data and the label constitute the training set of the sequence model. Further, the training set is used to train the sequence model, so that the output of the sequence model can be close to the corresponding extraction. problem, and then get the trained sequence model.

For most transaction scenarios, some "feature factors" need to be considered in model training, such as order status, logistics status, and the sequence model becomes unavailable. Therefore, this application adopts one of the deep learning CTR models for the transaction class. For example, the DIEN model uses the labeled data to perform supervised training on the DIEN model. After the training is completed, it is used as a “transaction” scenario, combining discrete orders and other information to predict user problems. The DIEN model is an end-to-end model. Predictive models for solving forecasting problems.

Optionally, the transaction information may include: order, logistics and other information, and the information may correspond to discrete features or continuous features, for example, status information is discrete features, and price information is continuous features. FIG. 6 is a schematic diagram of training a deep interest evolution network model provided by an embodiment of the present application; as shown in FIG. 6 , in the DIEN model training, historical interactive behavior sequences are used as one input, and price information is used as another input, such as discrete Features such as order status, logistics status or continuous features such as order price are used as another input, and the corresponding problem scenario is used as another input. Input the above input data into the DIEN model for training, and the trained DIEN model can be obtained. The output of the DIEN model can include the probability of clicking on the problem scene, and the probability is used to determine the probability of the problem encountered by the user. The output of the control model is in the top N values with the largest click probability, which can improve the problem. The accuracy of , where the likelihood measure is between 0 and 1, and N is a positive integer greater than 1.

Therefore, in the embodiment of the present application, a training data set can be constructed based on the historical conversation data with the intelligent customer service, so that the model can be trained based on the questions actually consulted by the user, and the accuracy of the model prediction can be improved.

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

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

Inputting the word segmentation sequence into the natural language model, and pre-training the natural language model with predicting the next sentence as a training target;

Taking the extracted questions as supervision signals, fine-tuning training is performed on the pre-trained natural language model according to the historical interaction behavior sequence.

In the embodiment of the present application, since the prediction of problems in the marketing category and account category can achieve high accuracy and precision only through the input of pure behavior sequences, the sequence model, such as a natural language model, can be used to process marketing, Account class problem prediction, but due to the particularity of the URL, there are many meaningless words in the historical interaction behavior sequence, such as "www", "myPage" are also considered as a word, so this application by the historical interaction behavior sequence After performing the page segmentation operation, a plurality of vectors are obtained, and then a training file is formed, and the natural language model is pre-trained by using the file. The natural language model can be an arbitrary model that calculates the probability of the next word segmentation in the sequence data, such as Bert model.

Specifically, the word segmentation operation is performed on the user's interactive behavior sequence within a certain period of time, that is, the page name and page URL are word-segmented. For example, the item of a certain page is word-segmented. The sequence of "keywords", that is, the keyword-based modeling processes a single item into more "semantic" information. After the word segmentation operation, the interactive behavior sequence becomes multiple vectors, and the vectors are used to pre-train the natural language model. , to obtain an unsupervised natural language model capable of inputting action sequences.

Exemplarily, FIG. 7 is a schematic diagram of pre-training a natural language model provided by an embodiment of the present application; as shown in FIG. 7 , a training data set is obtained, and the historical interaction behavior sequences in the training data set and the extracted questions are segmented. Operation, the word segmentation operation on the historical interaction behavior sequence may be to extract the word segmentation from the URL, and the word segmentation operation for the extracted question may be to extract the word segmentation from the corresponding question of the URL. "market app mtp app onion note weeex market app ikc redirect indexnewmycenter" corresponds to the word segmentation extracted from the URL, and the Chinese "why the account number is missing" corresponds to the word segmentation extracted from the corresponding question of the URL, and then the word segmentation sequence is obtained, Further, the word segmentation sequence is input into the natural language model, and the natural language model is pre-trained with predicting the next sentence as the training target, that is, the pre-trained natural language model is obtained by processing the input word segmentation sequence.

In this step, fine-tuning training can refer to finetune training. Finetune training is based on the pre-trained natural language model, using historical interaction behavior sequences and extraction problems to train a new model without completely retraining the natural language model, that is By using fewer iterations, a model with better effect can be obtained and the efficiency of model training can be improved.

Therefore, the embodiment of the present application can enable the natural language model to better process the behavior sequence through pre-training. During pre-training, the interaction sequence and the corresponding question can be used as model input together, so that the pre-trained model can be more Accurately extract the meaning of the interaction sequence and improve the pre-training effect.

Optionally, obtain the interactive behavior sequence of the user within the current preset time, including:

Through the stream computing framework, according to the user's interaction log within the current preset time, the stream data containing the interaction behavior sequence is generated;

The streaming data is filtered based on the rule logic library, and the filtered streaming data is put into a message queue, and the streaming data in the message queue is used for processing by the problem classification model.

In the embodiment of this application, streaming data is a key-value pair processed by a user log through a computing engine in a streaming computing framework. For example, if a user clicks a button in an application or visits a page, correspondingly, the A piece of log data is recorded in the log system, and further processed into a corresponding key-value pair by the computing engine, that is, the key formed by the user's ID number (Identity Document, ID) is the key and the sequence of URLs is the value. value pair.

Specifically, when the stream computing framework processes the log data, the implementation logic of the processed log data is packaged into a jar package for processing, and the stream data including the interactive behavior sequence is generated, and further, the stream data is It is submitted to the stream computing engine for execution in the form of a stream computing self-defined function, and further, an algorithm model Http online service interface is called, and the interface is used to filter the streaming data, that is, based on the dictionary, it is judged whether the stream data represents that the user has Consultation requirements, the dictionary includes multiple keywords and/or key pages, if there are consultation requirements, the streaming data is written into the message queue (MetaQ) in the form of messages; if there is no consultation requirement, the streaming data is discarded , and after obtaining the interactive behavior sequence of the user within the next preset time, the judgment is performed again by the above method.

Among them, judging whether the streaming data represents that the user has a consulting demand, that is, whether there is a consulting demand is determined by judging whether the number of keywords and/or key pages in the dictionary hit by the streaming data is greater than a preset threshold, an embodiment of the present application. The preset threshold is not specifically limited. For example, it can be determined that the number of keywords in the streaming data is greater than 5, and/or the number of key pages in the streaming data is greater than 1.

Optionally, the interactive behavior sequence may also include an interactive commodity sequence, a user button event on a page, and the like. For example, a user clicks a button event of a favorite operation on a certain page, the embodiment of the present application includes the interactive behavior sequence. The specific content is not limited, and it can be appropriately extended horizontally.

Therefore, the embodiment of the present application can generate streaming data based on a stream computing framework, use a rule logic library to filter the streaming data, filter out the streaming data that meets the requirements, and then input the streaming data that meets the requirements into the problem classification model. processing, reduce the interference of useless data, and improve the accuracy of problem classification.

Optionally, filter the streaming data based on the rule logic library, and put the filtered streaming data into a message queue, including:

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

Wherein, the dictionary includes a page dictionary and/or a page keyword dictionary;

The dictionary is determined by historical pages browsed by historical users and corresponding questions;

The rule is used to indicate whether the dictionary is used as a blacklist or a whitelist, and/or the number of pages in the page sequence browsed by the user that matches the dictionary is used to complete the screening operation on the page sequence.

In the embodiment of the present application, the dictionary is determined based on mutual information, and the mutual information refers to a useful information metric used to express the amount of information about keywords (keywords) contained in the problems encountered by the user, That is to measure the amount of information a keyword contributes to the problems encountered by users, which can be determined by the following formula:

Among them, x _i represents the ith keyword among all possible keywords extracted from the URL, _yi represents the ith problem among all possible problems encountered by the user, i represents the corresponding quantity, and I( Y, X) is used to measure the contribution of keywords to the problems encountered by users, and P(X= _xi , Y=y _i ) represents the probability of simultaneous occurrence of X= _xi and Y=y _i in the historical data. The historical data is the extracted problem corresponding to the URL and the URL obtained within a period of time, P(X= _xi ) represents the probability of X= _xi in the URL in the historical data, and P(Y=y _i ) represents the history The probability of occurrence of Y=y _i in the data, the above-mentioned probabilities can be obtained by calculating the corresponding data in the historical data, for example, P(X= _xi ) can be obtained by calculating the number of keywords extracted from the URL in the historical data and the historical data. All data may constitute the ratio of the corresponding number of word segmentation.

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

In the embodiment of this application, the keyword is split from the page name or URL. For example, if the page name is the refund details page, it can be split into "refund", "details", and "page" For these three keywords, further, the mutual information of the three keywords is calculated respectively, and the keyword whose mutual information is greater than the set threshold is the page keyword in the dictionary.

Among them, there is a corresponding page dictionary for the page name or URL, and there is a corresponding page keyword dictionary for the keywords in the page name or URL. If the dictionary only includes the page dictionary, when the streaming data is filtered, that is, Determine how many page names or URLs in the streaming data hit the corresponding page names or URLs in the dictionary; if the dictionary only includes a page keyword dictionary, when the streaming data is filtered, it is determined that the streaming data How many page names or keywords in the URL hit the corresponding page keywords in the dictionary in the data; if the dictionary includes a page dictionary and a page keyword dictionary, when filtering the streaming data, it can be performed based on preset conditions. Screening; the preset conditions include at least one of the following: the page name or URL in the streaming data hits the corresponding page name or URL in the page dictionary, and the keyword in the page name or URL in the streaming data hits The corresponding page keywords in the page keyword dictionary, that is, the streaming data can be filtered by determining that any one or all of the preset conditions are satisfied.

In this step, since the query rate per second that the Http online service interface of the algorithm model needs to undertake is extremely large, the present application screens the streaming data through a logic rule base. The rule logic base consists of rules, and the calculation of the rules depends on The following two types of dictionaries are implemented: a page dictionary and a page keyword dictionary. Specifically, the rules are used to indicate whether the dictionary is used as a blacklist or a whitelist, and the blacklist indicates the mutual information of pages and/or keywords in the dictionary. If it is smaller, the whitelist indicates that the mutual information of the pages and/or keywords in the dictionary is larger.

Optionally, the keywords can also be filtered, because too many keywords will increase the workload of streaming data filtering and affect the efficiency of the model. Therefore, it is proved by experiments that the range of keywords is reduced to Around 1000 is an ideal value.

Optionally, if the interactive behavior sequence includes the user's button event on the page, correspondingly, there is a corresponding button event dictionary for the button event, such as by judging the touch operation in the page browsed by the user and the button event in the dictionary. The filtering operation of the streaming data is completed after matching. Similarly, if the interactive behavior sequence also includes other content, such as an interactive commodity sequence, there is also a corresponding dictionary, and the dictionary can be used to filter the streaming data. This embodiment of the present application The number of rules contained in the rule logic base is not specifically limited, and the contents contained in the interactive behavior sequence have corresponding dictionaries and rules for performing screening operations.

Therefore, after filtering the streaming data through the rules in the dictionary and the rule logic library, the embodiment of the present application can write important data into the message queue, reduce the amount of tasks undertaken by the message queue, reduce the back pressure of the stream computing engine, or reduce the important loss of data.

Optionally, an embodiment of the present application also provides a specific solution for determining a dictionary based on mutual information. FIG. 8 provides a flowchart for determining a dictionary based on mutual information provided by an embodiment of the present application; the process includes:

S801. Acquire historical conversation data of multiple historical users and intelligent customer service, and extract corresponding questions according to the historical conversation data to obtain a question set.

In this step, the set of questions is the conversational communication between historical users and intelligent customer service, and then the problems encountered by users extracted in the background, for each historical user and intelligent customer service for conversational communication, and at least one question can be extracted from each conversational communication , the problem set is used to express various problems that users may encounter.

S802. Determine a page set and a page keyword set browsed by the multiple historical users according to the historical browsing page sequence of the multiple historical users before the historical session data.

In this step, the historical browsing page sequence before the historical session data corresponding to the multiple historical users is decomposed, and then the page set and the page keyword set browsed by the multiple historical users are determined; the page set includes at least one of the historical users. page and at least one page keyword.

S803. For any question in the question set, calculate the mutual information between each page in the page set and the question, and the mutual information between each page keyword in the page keyword set and the question, and calculate to obtain The mutual information selects the target page and target page keyword corresponding to the question from the page set and the page keyword set.

In the embodiment of the present application, the determined target page contains keywords that can be used to predict user problems, and the target page is a page that includes useful information, that is, pages corresponding to the user's regular browsing behavior are not target pages, such as those frequently visited by users. Pages such as "Home", "My", etc., which cannot be considered as target pages.

Specifically, after calculating the mutual information between each page and the question in the page set by using formula (1), the mutual information is sorted in order of size. Further, formula (1) is used to calculate the key of each page in the page keyword set. The mutual information between the word and the question is also sorted by the calculated mutual information, and then the pages and page keywords that are both ranked in the top N are selected, and then the target page and the target page keyword corresponding to the question are determined.

S804 , fuse the target pages corresponding to each question in the question set to obtain a page dictionary, and fuse the target page keywords corresponding to each question to obtain a page keyword dictionary.

Exemplarily, it is assumed that there are 100 corresponding questions extracted from historical session data, 1,000 pages determined from the historical browsing page sequence, and 10,000 page keywords. Further, each page pair is calculated based on mutual information. The contribution of each extracted question, and the contribution of each page keyword to each extracted question, after the calculation is completed, for each question, find out the top 1000 keywords and the top 100 pages corresponding to the question, 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 keys corresponding to all the questions are Words are fused to get the page keyword dictionary.

It should be noted that the embodiments of the present application do not specifically limit the number of target pages and target page keywords corresponding to each question, nor do they specifically limit the number of extracting corresponding questions, and the above are only examples.

Therefore, the embodiment of the present application can extract the problems encountered by the user according to the actual historical session data of the user, and then determine the contribution of each page and keyword to the problem encountered by the user in combination with the page sequence that the user browsed before consulting the problem, so as to determine the contribution of each page and keyword to the problem encountered by the user based on mutual information Accurately determine the page dictionary and page keyword dictionary in the dictionary, improve the accuracy of determining the dictionary, and reduce the interference of useless data.

With reference to the above-mentioned embodiments, FIG. 9 is an online calculation flowchart of a problem prediction method provided by the embodiment of the application. As shown in FIG. 9 , after the stream computing framework stores tasks (streaming data) in the message middleware, the Tasks are processed by filters, vectorization, tagging, sorting modules and process pools in turn to predict the problems encountered by users. Among them, the filter processes real-time pages such as active pages based on the filtering logic (rule logic library). Keyword items or key pages and rule parameters; vectorization is processed based on the representation learning model, that is, in the embedding stage, the user's serialized page browsing path is encoded into a two-dimensional vector, and the historical interaction behavior sequence and extraction problem are used. The model is fine-tuned and trained, and the historical interaction behavior sequence and the extracted questions are processed by the basic tri-convolutional neural network model (question classification model), and labelled, that is, classified; the sorting module is used for questions/knowledge (problems encountered by users); the process pool is used for buffering/triggering, ie triggering push messages to users.

The tasks acquired by the stream computing framework come from the log service (Log Service, SLS). The stream computing framework includes a service module, a refined operation module, and an application configuration center. During the message consumption process of the stream computing framework, the message consumption output The data type can include voice, picture, text or video, etc. Further, based on the output data type, select the corresponding recommendation form to recommend the data to the user, and the recommendation form can include intelligent outbound call, message, message/SMS , customer service, etc.; for example, the solution can be intelligently called out to users in the form of voice through phone calls.

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

Table 1

Among them, the problem based on the output of the sorting module is not consistent with the card content that is finally pushed to the client. Therefore, it is necessary to maintain a set of pushed card content list, including the mapping relationship between the problems encountered by the user and the card ID, so that when the problem is predicted , the card ID corresponding to the problem ID can be found through the mapping table, and then the card content of the corresponding card ID can be called in the background and pushed to the client.

Optionally, the method provided in this application may not be coupled with any kind of push system. When the output problem is determined, the push system sends a message that triggers the push, which can be in any other form, such as cards, copywriting, etc. The above table can also be used. Other designs are added according to actual needs, which are not specifically limited in this embodiment of the present application.

In this embodiment of the present application, relevant indicators can also be displayed to operators, so that operators can view the effect of problem prediction, and further, operators can formulate actions for refined operations; the relevant indicators may include the number of users sent, the number of users successfully sent The number of users, the number of users reached, the number of effective users, the number of active service sessions, the active service resolution rate, the active service ratio, the reach rate, the effective rate and the click rate, etc.

Optionally, the amount of sending users refers to the amount of push users received by the push system; the amount of users that are successfully sent refers to the number of successful push solutions, that is, users have received push content; the reach users The amount refers to the number of users who open the pushed message and see the push card, in which the message becomes read; the effective user amount refers to the number of valid actions that the user performs on the basis of seeing the card, such as clicking , or reply information, etc.; the active service session volume refers to the number of session services generated on the smart customer service page after the user opens the push message and sees the push card; the active service resolution rate can be the number of successful solutions and possible solutions. The ratio of the number of successful solutions can be the number of active service sessions minus the number of unsuccessful solutions, and the number of unsuccessful solutions can include: the number of online transfers, the number of calls transferred, the number of unsatisfactory sessions, the number of unsatisfactory sessions, the last recommendation The number of clicked sessions, the last unanswered session, the amount of direct connection labor, etc., the number of possible solutions can be the amount of active service sessions minus the amount of direct connection labor; the active service ratio refers to the number of active service sessions and the The ratio of the total number of service sessions (including active and passive sessions) in the application; the reach rate refers to the ratio of the number of reached users to the number of successfully sent users; the effective rate refers to the number of effective users and the number of touch The ratio of the number of users reached; the click rate refers to the ratio of the number of users who clicked to the number of users reached.

Optionally, the operation action may include: setting the type and quantity of the recommended questions, adding, deleting, checking, and modifying the recommended questions, controlling the fatigue degree of the push crowd, editing the push template/question copy, and screening the push crowd, etc.

Exemplarily, during the screening of push groups, messages can be filtered to reach insensitive groups, negative groups, or blacklists, etc., that is, for insensitive groups and negative groups, messages can be not pushed or less pushed, and then Improve user experience.

Optionally, the timing of the message push is triggered by the user's browsing behavior on the application page. At this time, the application is in an open state, and the user is reached by means of message push, which can ensure a higher reach rate and efficiency.

In another possible implementation manner, the behavior log data can be directly monitored by means of a distributed task system and handwritten programming logic, and then the problem prediction method shown in FIG. 3 can be performed. The method is not specifically limited, and other frameworks may be used instead of the flow computing framework in the embodiments of the present application, as long as the same active problem prediction as the present application can be achieved, it is within the protection scope of the present application.

Therefore, the problem prediction method provided in the embodiment of the present application has strong push timeliness, can be triggered by the user's browsing behavior, and produces the user's problem prediction result in seconds, thereby improving a higher reach rate and efficiency; Instead of relying on crowd labels, it relies on multi-layer and multiple combined in-depth prediction algorithm models to make judgments and improve the accuracy of the prediction model. The input data used in this application are the user's dynamic behavior data and dynamic order status data. The output For the problem of intelligent service scenarios, it can more accurately capture real-time user demands, thereby improving the user experience of intelligent services.

FIG. 10 is a schematic flowchart of another problem prediction method provided by the embodiment of the present application. As shown in FIG. 10 , a solution can be pushed to the seller according to the buyer's problem, and the method includes:

S1001. Acquire an interactive behavior sequence of at least one buyer corresponding to the store within a current preset time; wherein, the interactive behavior sequence includes a browsed page sequence.

In the embodiment of the present application, at least one buyer corresponding to a store may refer to a buyer who has a purchase record, a buyer who has a comment record, or a buyer who has a consultation record in the store.

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

S1002. According to the interactive behavior sequence, predict the types of problems encountered by each buyer based on a problem classification model, and predict the problems encountered by each buyer based on a problem prediction model corresponding to the type.

In this embodiment of the present application, the problem classification model obtains the types of problems encountered by each buyer, and the classification of the types can be classified according to the method described in step 302, which is not repeated here, or can be classified according to the buyer For example, when the number of buyers is large, the model with fast calculation speed can be used, and when the number of buyers is small, the model with average calculation speed can be used to reduce the calculation amount of the default system. , the classification of the types is not specifically limited in the embodiment of the present application.

S1003. Determine a solution to be pushed to the seller corresponding to the store according to the predicted problem.

In this step, the solution to be pushed to the seller corresponding to the store is determined, and the corresponding solution can be selected and recommended to the seller based on the type and size of the store, such as a larger store, the buyer with a higher priority can be selected The corresponding solution is pushed to the store first. The buyer with high priority can refer to member customers or buyers with good shopping credit, etc. For smaller stores, you can choose the solution corresponding to all buyers. Send to the store.

It should be noted that the push form and search method of the solution are similar to step 303 in the above-mentioned embodiment, which will not be repeated here, and reference may be made to the description of step 303 . Optionally, for the same problem, the solution for buyers and the solution for sellers can be different. For example, for sellers, if there are multiple buyers who have refund problems, they can push a copy to the seller. : Multiple buyers may have issues with refunds, please click here for solutions.

Therefore, in the embodiment of the present application, the acquired page sequence of at least one buyer in the current preset time can be used; the type of the problem encountered by each buyer can be predicted based on the problem classification model, and the problem prediction model corresponding to the type can be used. , proactively predict the problems encountered by each buyer; further, according to the predicted problems, determine the solution that is pushed to the seller corresponding to the store, and improve the efficiency and accuracy of the seller's solution to the user's problem.

Optionally, according to the interactive behavior sequence, predict the types of problems encountered by each buyer based on the problem classification model, and predict the problems encountered by each buyer based on the problem prediction model corresponding to the type, including:

For any buyer, input the feature vector corresponding to the interactive behavior sequence into the problem classification model to determine the type of problem encountered by the buyer;

If the type is a marketing type or an account type, processing the interactive behavior sequence of the buyer based on the sequence model to obtain the corresponding question;

If the type is a transaction type, the interaction behavior sequence of the buyer and the transaction information corresponding to the buyer are processed based on the deep interest evolution network model to obtain corresponding questions.

Optionally, historical conversation data between historical sellers and intelligent customer service can also be obtained, and corresponding questions can be extracted according to the historical conversation data;

According to the historical interaction behavior sequence of the historical seller corresponding to the historical buyer before the historical session data and the extracted questions, construct a training data set corresponding to the sequence model, and train the sequence model; and/or,

According to the historical interaction behavior sequence, transaction information and extracted questions of historical buyers of the historical seller before the historical session data, a training data set corresponding to the deep interest evolution network model is constructed, and the deep interest evolution network model is to train.

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

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

Inputting the word segmentation sequence into the natural language model, and pre-training the natural language model with predicting the next sentence as a training target;

Taking the extracted questions as supervision signals, fine-tuning training is performed on the pre-trained natural language model according to the historical interaction behavior sequence.

Optionally, obtain the interactive behavior sequence of at least one buyer corresponding to the store within the current preset time, including:

Through the stream computing framework, according to the interaction log of at least one buyer corresponding to the store within the current preset time, the stream data including the sequence of interaction behaviors is generated;

The streaming data is filtered based on the rule logic library, and the filtered streaming data is put into a message queue, and the streaming data in the message queue is used for processing by the problem classification model.

Optionally, filter the streaming data based on the rule logic library, and put the filtered streaming data into a message queue, including:

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

Wherein, the dictionary includes a page dictionary and/or a page keyword dictionary;

The dictionary is determined by historical pages browsed by historical buyers and corresponding questions;

The rule is used to indicate whether the dictionary is used as a blacklist or a whitelist, and/or the number of pages in the page sequence browsed by the buyer that matches the dictionary is used to complete the screening operation on the page sequence.

Optionally, also include:

Obtaining historical conversation data of multiple historical sellers and intelligent customer service, extracting corresponding questions according to the historical conversation data, and obtaining a set of questions;

Determine the page set and page keyword set browsed by the historical buyer according to the historical browsing page sequence of the historical buyer before the historical session data of the plurality of historical sellers;

For any question in the question set, calculate the mutual information between each page in the page set and the question, and the mutual information between each page keyword in the page keyword set and the question, and calculate the mutual information obtained. The information selects the target page and target page keyword corresponding to the question from the page set and the page keyword set;

The target pages corresponding to each question in the question set are fused to obtain a page dictionary, and the target page keywords corresponding to each question are fused to obtain a page keyword dictionary.

The implementation principle and technical effect of the foregoing embodiment are similar to those of the foregoing embodiment of the subordinate solution corresponding to FIG. 3 , and details are not described herein again.

Corresponding to the above problem prediction method, an embodiment of the present application provides a problem prediction apparatus. FIG. 11 is a schematic structural diagram of a problem prediction apparatus provided by an embodiment of the present application, and the apparatus includes:

The first obtaining module 1101 is used to obtain the interactive behavior sequence of the user within the current preset time; wherein, the interactive behavior sequence includes the browsed page sequence;

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

The first determination module 1103 is 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 can be used to implement the technical solutions of the embodiments shown in FIG. 1 to FIG. 9 , and the implementation principle and technical effect thereof are similar, and are not described again in this embodiment.

Exemplarily, an embodiment of the present application further provides an apparatus for predicting a problem. FIG. 12 is a schematic structural diagram of another apparatus for predicting a problem provided by an embodiment of the present application, and the apparatus includes:

The second acquiring module 1201 is configured to acquire the interactive behavior sequence of at least one buyer corresponding to the store within the current preset time; wherein, the interactive behavior sequence includes the browsed page sequence;

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

The second determination module 1203 is configured to determine a solution to be pushed to the seller corresponding to the store according to the predicted problem.

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

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 this embodiment may include:

at least one processor 1301; and

a memory 1302 in communication with the at least one processor;

Wherein, the memory 1302 stores instructions that can be executed by the at least one processor 1301, and the instructions are executed by the at least one processor 1301, so that the electronic device executes any of the foregoing embodiments. method.

Optionally, the memory 1302 may be independent or integrated with the processor 1301 .

For the implementation principle and technical effect of the electronic device provided in this embodiment, reference may be made to the foregoing embodiments, and details are not repeated here.

Embodiments of the present application further provide a computer-readable storage medium, where computer-executable instructions are stored in the computer-readable storage medium, and when a processor executes the computer-executable instructions, the method described in any of the foregoing embodiments is implemented.

Embodiments of the present application further provide a computer program product, including a computer program, which implements the method described in any of the foregoing embodiments when the computer program is executed by a processor.

In the technical solution of this application, the collection, storage, use, processing, transmission, provision and disclosure of user data and other information involved are in compliance with relevant laws and regulations, and do not violate public order and good customs.

In the several embodiments provided in this application, it should be understood that the disclosed apparatus and method may be implemented in other manners. For example, the device embodiments described above are only illustrative. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated. to another system, or some features can be ignored, or not implemented.

The above-mentioned integrated modules implemented in the form of software functional modules may be stored in a computer-readable storage medium. The above-mentioned software function modules are stored in a storage medium, and include several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor to execute some steps of the methods described in the various embodiments of the present application.

It should be understood that the above-mentioned processor may be a central processing unit (Central Processing Unit, CPU for short), or other general-purpose processors, digital signal processors (Digital Signal Processor, DSP for short), application specific integrated circuit (Application Specific Integrated Circuit, Referred to as ASIC) and so on. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of the method disclosed in conjunction with the application can be directly embodied as executed by a hardware processor, or executed by a combination of hardware and software modules in the processor. The memory may include high-speed RAM memory, and may also include non-volatile storage NVM, such as at least one magnetic disk memory, and may also be a U disk, a removable hard disk, a read-only memory, a magnetic disk or an optical disk, and the like.

The above-mentioned storage medium may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read only memory (EEPROM), erasable Except programmable read only memory (EPROM), programmable read only memory (PROM), read only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. A storage medium can be any available medium that can be accessed by a general purpose or special purpose computer.

An exemplary storage medium is coupled to the processor, such that the processor can read information from, and write information to, the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and the storage medium may be located in application specific integrated circuits (Application Specific Integrated Circuits, ASIC for short). Of course, the processor and the storage medium may also exist in the electronic device or the host device as discrete components.

It should be noted that, herein, the terms "comprising", "comprising" or any other variation thereof are intended to encompass non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements, It also includes other elements not expressly listed or inherent to such a process, method, article or apparatus. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in a process, method, article or apparatus that includes the element.

The above-mentioned serial numbers of the embodiments of the present application are only for description, and do not represent the advantages or disadvantages of the embodiments.

From the description of the above embodiments, those skilled in the art can clearly understand that the method of the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is better implementation. Based on this understanding, the technical solution of the present application can be embodied in the form of a software product in essence or in a part that contributes to the prior art, and the computer software product is stored in a storage medium (such as ROM/RAM, magnetic disk, CD-ROM), including several instructions to make a terminal device (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the methods described in the various embodiments of this application.

The above are only the preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made by using the contents of the description and drawings of the present application, or directly or indirectly applied in other related technical fields , are similarly included within the scope of patent protection of this application.

Claims (11)

1. a problem prediction method, is characterized in that, described method comprises: Acquire the interactive behavior sequence of the user within the current preset time; wherein, the interactive behavior sequence includes the browsed page sequence; According to the interactive behavior sequence, the type of the problem encountered by the user is predicted based on the problem classification model, and the problem encountered by the user is predicted based on the problem prediction model corresponding to the type; Based on the predicted problem, a solution to push to the user is determined. 2. The method according to claim 1, wherein, according to the interactive behavior sequence, the type of the problem encountered by the user is predicted based on a problem classification model, and the problem prediction model corresponding to the type is used to predict the problem. problems encountered by users, including: Input the feature vector corresponding to the interactive behavior sequence into the problem classification model to determine the type of problem encountered by the user; If the type is a marketing type or an account type, the user's interactive behavior sequence is processed based on the sequence model to obtain the corresponding question; If the type is a transaction type, the interaction behavior sequence of the user and the transaction information corresponding to the user are processed based on the deep interest evolution network model to obtain the corresponding problem. 3. The method of claim 2, further comprising: Obtain historical conversation data between historical users and intelligent customer service, and extract corresponding questions according to the historical conversation data; According to the historical interaction behavior sequence of the historical user before the historical session data and the extracted questions, a training data set corresponding to the sequence model is constructed, and the sequence model is trained; and/or, According to the historical interaction behavior sequence, transaction information and extracted questions of the historical user before the historical session data, a training data set corresponding to the deep interest evolution network model is constructed, and the deep interest evolution network model is trained. 4. The method according to claim 3, wherein the sequence model is a natural language model for processing sequences; training the sequence model, comprising: Perform a word segmentation operation on the historical interactive behavior sequence and the extracted questions in the training data set to obtain a word segmentation sequence; Inputting the word segmentation sequence into the natural language model, and pre-training the natural language model with predicting the next sentence as a training target; Taking the extracted questions as supervision signals, fine-tuning training is performed on the pre-trained natural language model according to the historical interaction sequence. 5. The method according to any one of claims 1-4, characterized in that, acquiring the interactive behavior sequence of the user within the current preset time, comprising: Through the stream computing framework, according to the user's interaction log within the current preset time, the stream data containing the interaction behavior sequence is generated; The streaming data is filtered based on the rule logic library, and the filtered streaming data is put into a message queue, and the streaming data in the message queue is used for processing by the problem classification model. 6. The method according to claim 5, wherein the streaming data is screened based on a rule logic library, and the filtered streaming data is put into a message queue, comprising: Screening the streaming data according to the dictionary and the rules in the rule logic library; Wherein, the dictionary includes a page dictionary and/or a page keyword dictionary; The dictionary is determined by historical pages browsed by the user and corresponding questions. 7. The method of claim 6, further comprising: Obtaining historical conversation data of multiple historical users and intelligent customer service, extracting corresponding questions according to the historical conversation data, and obtaining a set of questions; determining a page set and a page keyword set browsed by the plurality of historical users according to the sequence of historical pages viewed by the plurality of historical users before the historical session data; For any question in the question set, calculate the mutual information between each page in the page set and the question, and the mutual information between each page keyword in the page keyword set and the question, and calculate the mutual information obtained. The information selects the target page and target page keyword corresponding to the question from the page set and the page keyword set; The target pages corresponding to each question in the question set are fused to obtain a page dictionary, and the target page keywords corresponding to each question are fused to obtain a page keyword dictionary. 8. A problem prediction method, comprising: Obtain the interactive behavior sequence of at least one buyer corresponding to the store within the current preset time; wherein, the interactive behavior sequence includes the browsed page sequence; According to the interactive behavior sequence, the type of the problem encountered by each buyer is predicted based on the problem classification model, and the problem encountered by each buyer is predicted based on the problem prediction model corresponding to the type; According to the predicted problem, a solution to be pushed to the seller corresponding to the store is determined. 9. An electronic device, characterized in that, 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, and the instructions are executed 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, characterized in that, computer-executable instructions are stored in the computer-readable storage medium, and when a processor executes the computer-executable instructions, the computer-executable instructions according to any one of claims 1-8 are implemented. method described. 11. A computer program product, comprising a computer program, characterized in that, when the computer program is executed by a processor, the method according to any one of claims 1-8 is implemented.

Images