CN112885337A

CN112885337A - Data processing method, device, equipment and storage medium

Info

Publication number: CN112885337A
Application number: CN202110130657.XA
Authority: CN
Inventors: 杨海军; 徐倩; 杨强
Original assignee: WeBank Co Ltd
Current assignee: WeBank Co Ltd
Priority date: 2021-01-29
Filing date: 2021-01-29
Publication date: 2021-06-01

Abstract

The embodiment of the invention provides a data processing method, a device, equipment and a storage medium, wherein the method comprises the following steps: the method comprises the steps of obtaining model parameters respectively sent by a plurality of clients, wherein the model parameters sent by each client are model parameters obtained after the client trains a global model according to local training data, and the local training data are data obtained in the intelligent seat service of the corresponding client; aggregating the model parameters sent by the plurality of clients to obtain an updated global model; and processing the input information according to the updated global model to obtain a corresponding auxiliary result, wherein the auxiliary result is used for assisting to determine a response reply corresponding to the input information. The global model is obtained by performing federal learning through local training data of a plurality of clients, so that the accuracy and the integrity of the model are higher, the accuracy of the obtained auxiliary result is higher, and the response assistance of the agent assistant to the agent is facilitated.

Description

Data processing method, device, equipment and storage medium

Technical Field

The present invention relates to the field of artificial intelligence technologies, and in particular, to a data processing method, apparatus, device, and storage medium.

Background

The agent assistant is a system tool which is based on voice recognition, semantic analysis and the like and provides agent response auxiliary capacity for the agent personnel, and can effectively help the agent personnel to improve the working efficiency, the service level and the like.

The current seat assistant mainly realizes the service of seat response assistance through model training in advance. Taking voice recognition as an example, voice call needs real-time voice recognition and conversion into characters, and the seat assistant provides recommended speech for seat personnel according to the result of voice recognition. Therefore, it is necessary to acquire the voice annotation data as a training sample in advance to train the voice recognition model. However, the effect of the speech recognition model can be better only by requiring a large amount of labeled data of rich scenes, and the data of each system are not intercommunicated at present, so that the labeled data capable of covering a large amount of scenes are difficult to obtain for model training, and the accuracy of the model is low.

Therefore, the existing seat assistant is difficult to cover a large number of scenes due to the used model, so that the accuracy of the response assistance provided by the seat assistant is low.

Disclosure of Invention

The invention mainly aims to provide a data processing method, a data processing device, data processing equipment and a data processing storage medium, and aims to solve the problem that the response assistance provided by the existing seat assistant is low in accuracy.

To achieve the above object, the present invention provides a data processing method, including:

the method comprises the steps of obtaining model parameters respectively sent by a plurality of clients, wherein the model parameters sent by each client are model parameters obtained after the client trains a global model according to local training data, and the local training data are data obtained in the intelligent seat service of the corresponding client;

aggregating the model parameters sent by the plurality of clients to obtain an updated global model;

and processing the input information according to the updated global model to obtain a corresponding auxiliary result, wherein the auxiliary result is used for assisting to determine a response reply corresponding to the input information.

In a possible implementation manner, the processing the input information according to the updated global model to obtain a corresponding auxiliary result includes:

processing the input information according to the updated global model to obtain an intermediate result;

and determining the corresponding auxiliary result according to the intermediate result.

In one possible implementation, the global model is a classification model, the local training data includes a plurality of groups of first sample data, each group of first sample data includes first sample information and a sample classification label of the first sample information; the processing the input information according to the updated global model to obtain an intermediate result includes:

and processing the input information according to the classification model to obtain a classification label corresponding to the input information.

In a possible implementation, the determining the corresponding auxiliary result according to the intermediate result includes:

and determining a classification result corresponding to the classification label in a first database according to the classification label corresponding to the input information.

In a possible implementation manner, the input information is text information, the global model is a semantic understanding model, the local training data includes a plurality of sets of second sample data, and each set of second sample data includes sample text information and a sample semantic tag of the sample text information; the processing the input information according to the updated global model to obtain an intermediate result includes:

and processing the text information according to the semantic understanding model to obtain a semantic label corresponding to the text information.

and determining at least one search answer corresponding to the semantic label in a second database according to the semantic label corresponding to the text information.

In a possible implementation manner, the input information is speech information, the global model is a speech recognition model, the local training data includes a plurality of sets of third sample data, and each set of third sample data includes sample speech information and a sample speech recognition result of the sample speech information; the processing the input information according to the updated global model to obtain an intermediate result includes:

and processing the voice information according to the voice recognition model to obtain a voice recognition result corresponding to the voice information.

and determining at least one response text corresponding to the voice recognition result in a third database according to the voice recognition result corresponding to the voice information.

In a possible implementation manner, the input information is user information, the global model is a user portrait model, the local training data includes a plurality of sets of fourth sample data, and each set of fourth sample data includes sample user information and a sample user identifier of the sample user information; the processing the input information according to the updated global model to obtain an intermediate result includes:

and processing the user information according to the user portrait model to obtain a user identifier corresponding to the user information.

and determining a user label corresponding to the user identification in a fourth database according to the user identification corresponding to the user information.

The present invention also provides a data processing apparatus, the apparatus comprising:

the system comprises an acquisition module, a storage module and a processing module, wherein the acquisition module is used for acquiring model parameters respectively sent by a plurality of clients, the model parameters sent by each client are model parameters obtained after the client trains a global model according to local training data, and the local training data are data obtained in the intelligent seat service of the corresponding client;

the training module is used for carrying out aggregation processing on the model parameters sent by the plurality of clients to obtain an updated global model;

and the processing module is used for processing the input information according to the updated global model to obtain a corresponding auxiliary result, and the auxiliary result is used for assisting in determining a response reply corresponding to the input information.

In a possible implementation, the processing module is specifically configured to:

In one possible implementation, the global model is a classification model, the local training data includes a plurality of groups of first sample data, each group of first sample data includes first sample information and a sample classification label of the first sample information; the processing module is specifically configured to:

In a possible implementation manner, the input information is text information, the global model is a semantic understanding model, the local training data includes a plurality of sets of second sample data, and each set of second sample data includes sample text information and a sample semantic tag of the sample text information; the processing module is specifically configured to:

In a possible implementation manner, the input information is speech information, the global model is a speech recognition model, the local training data includes a plurality of sets of third sample data, and each set of third sample data includes sample speech information and a sample speech recognition result of the sample speech information; the processing module is specifically configured to:

In a possible implementation manner, the input information is user information, the global model is a user portrait model, the local training data includes a plurality of sets of fourth sample data, and each set of fourth sample data includes sample user information and a sample user identifier of the sample user information; the processing module is specifically configured to:

The present invention also provides a data processing apparatus, comprising: memory, a processor and a data processing program stored on the memory and executable on the processor, which when executed by the processor implements the steps of the data processing method according to any of the preceding claims.

The invention also provides a computer readable storage medium having stored thereon a data processing program which, when executed by a processor, implements the steps of the data processing method as claimed in any one of the preceding claims.

The invention also provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the data processing method according to any one of the preceding claims.

According to the data processing method, the data processing device, the data processing equipment and the storage medium, firstly, model parameters respectively sent by a plurality of clients are obtained, wherein the model parameters sent by each client are model parameters obtained after the client trains a global model according to local training data, and the local training data are data obtained in intelligent seat service of the corresponding client; then, carrying out aggregation processing on the model parameters sent by the plurality of clients to obtain an updated global model; and processing the input information according to the updated global model to obtain a corresponding auxiliary result, wherein the auxiliary result is used for assisting to determine a response reply corresponding to the input information. In the scheme of the embodiment of the invention, as the global model is obtained by performing federal learning on the respective local training data of the plurality of clients, and the sample data sources are wider and more comprehensive, the common global model obtained by training the sample data provided by the plurality of clients is higher in accuracy and integrity of the model and higher in accuracy of the obtained auxiliary result on the premise of not damaging the privacy of each client, and is more beneficial to response assistance of the assistant to the assistant.

Drawings

Fig. 1 is a schematic view of an application scenario provided in an embodiment of the present invention;

fig. 2 is a schematic diagram of a system architecture of an agent assistant according to an embodiment of the present invention;

fig. 3 is a schematic flow chart of a data processing method according to an embodiment of the present invention;

FIG. 4 is a system architecture diagram of federated learning model training provided by an embodiment of the present invention;

FIG. 5a is a schematic diagram of horizontal federated learning model training provided by an embodiment of the present invention;

FIG. 5b is a schematic diagram of longitudinal federated learning model training provided by an embodiment of the present invention;

FIG. 6 is a first schematic diagram illustrating an assistant result of an assistant according to an embodiment of the present invention;

FIG. 7 is a second schematic diagram illustrating assistance results of an agent assistant according to an embodiment of the present invention;

fig. 8 is a third schematic diagram illustrating an assistance result of the agent assistant according to the embodiment of the present invention;

FIG. 9 is a fourth schematic diagram illustrating assistance results of an agent assistant according to an embodiment of the present invention;

FIG. 10 is a block diagram of a data processing apparatus according to an embodiment of the present invention;

fig. 11 is a schematic structural diagram of a data processing device according to an embodiment of the present invention.

The implementation, functional features and advantages of the objects of the present invention will be further explained with reference to the accompanying drawings.

Detailed Description

Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

First, an application scenario applicable to the embodiment of the present invention is described.

Fig. 1 is a schematic view of an application scenario provided by an embodiment of the present invention, as shown in fig. 1, including a user 11, an agent person 12, and an agent assistant 13. The seat staff 12 can communicate with the user 11, and the seat assistant 13 can acquire the communication information of the user 11 and the seat staff 12 during communication, perform corresponding processing according to the communication information of the user 11 and the seat staff 12 during communication, and display the communication information to the seat staff 12 to help the seat staff 12 to reply with the user 11.

Fig. 2 is a schematic diagram of a system architecture of an assistant according to an embodiment of the present invention, and as shown in fig. 2, the assistant may provide all-around services for an agent, for example, the services may mainly include providing user portrayal, intelligent recommendation, answer search, and intelligent quality inspection.

User portrait: including basic attribute information of the user, an analysis tag, and the like. The agent assistant can establish an all-directional user portrait by automatically analyzing service system data and the like. When the seat personnel converse with the user, the user portrait information is displayed on the conversation page in real time, the seat personnel is released from heavy labor of manually arranging the user information, and the information displayed by the user portrait module is richer and more timely.

Intelligently recommending: the questions provided by the user are analyzed in real time, relevant answers are searched from the knowledge base and ranked from big to small according to the relevance, and the most relevant answers are recommended to an agent service page for the agent to reply by reference, so that the accuracy and efficiency of agent reply are improved, and the user experience is improved.

Answer searching: when the seat personnel want to search the answer of a certain question, the question which the seat personnel want to know can be input by using answer search, the answer search searches relevant questions from a knowledge base and sorts the questions according to the relevance from large to small, and then all answer sequences are output to the seat personnel for reference.

Intelligent quality inspection: and performing real-time quality inspection on the agent conversation text according to the pre-configured quality inspection item requirements, feeding back a quality inspection result to the agent, prompting the attention items in the agent conversation in time, and supervising and standardizing the agent behavior. Quality control of violence-induced quality control items can detect violent actions such as expressors and terrorists in conversations and send out warnings.

When the agent assistant provides the service, the agent assistant needs to acquire relevant information. In the embodiment of the present invention, the channel for the agent assistant to obtain the relevant information may include a voice customer service channel, such as telephone voice, online audio, and the like. The channels for acquiring the related information may also include text customer service channels, such as mobile phone APP, web pages, and the like.

Current agent assistants need to provide the services described above through pre-trained models. The involved models include a classification system and a classification model, a speech recognition model, a semantic understanding model, a user portrait model and the like.

When the classification system and the classification model are trained, the obtained training sample usually includes some information to be processed and labeled data corresponding to the information to be processed, and the labeled data is the classification to which the information to be processed belongs. Taking the banking industry as an example, the classification of the banking scene includes different classifications of borrowing, repayment, deposit and the like. The information to be processed may be a voice conversation between the user and the agent person, or may be text information input by the agent person, or the like. According to a certain bank behavior example, when a classification system and a classification model need to be trained, historical information to be processed of the bank can be obtained and marked. For example, the user has previously asked "how did your bank's interest on borrowing? And determining the classification of the voice information as "borrow", so that the voice information can be used as a training sample for training a classification system and a classification model after being labeled.

Aiming at the voice recognition model, when training, corresponding voice information is obtained at first, and a corresponding voice recognition result is marked to form a training sample for training. Aiming at the semantic understanding model, corresponding text information is firstly obtained, corresponding semantic recognition results are marked out, and a training sample is formed for training.

For the user portrait model, during training, identity information of a user can be obtained, and then other information of the user under the identity information can be obtained. Taking banking as an example, other information of the user may include consumption condition, financial condition, etc. of the user, so as to realize training of the user portrait model.

The agent assistant mainly realizes the agent response auxiliary service through a pre-trained model, so the accuracy of the model is very important for the agent assistant. The models trained in the existing seat assistant are obtained by training through data provided by a single data party, information sharing with other data is not performed, the provided data are single and limited, and a data isolated island is formed, so that the accuracy of the models obtained by training is low, the quality of service provided by the seat assistant is poor, and the accuracy of response assistance is low. The concrete aspects are as follows:

first, the intention classification system of the assistant is not sound enough, and the intention classification accuracy is not high enough. Taking the financial industry as an example, an intelligent question-answering classification system in a bank scene comprises dozens of major categories such as borrowing, repayment, deposit, interest, financing and the like, and hundreds of minor categories at the next layer. The knowledge base of each bank may only contain a part of classified contents when being edited, and the whole category is difficult to be covered. Therefore, the classification system is divided and is difficult to completely express; the data under each classification is also segmented, and a classification system and a classification model with higher accuracy are difficult to train by using more and more complete data.

Secondly, the semantic understanding ability of the assistant is not strong enough, that is, the understanding accuracy of multiple descriptions of the same semantic is not high enough. For example, in the banking industry, there are other different questions about "how to calculate interest on loan", such as "calculating method of borrowing money", "how much interest should be paid after borrowing 1 ten thousand money for 7 days", and so on. When a knowledge base is configured, each bank often enumerates only a few known problem sets which are split from each other, and the problem sets cannot be combined to train a semantic recognition model, so that the recall rate and the accuracy rate of problem semantic matching are low.

Thirdly, the voice recognition effect of the seat assistant is not good enough, namely, the voice call needs to be subjected to real-time voice recognition and converted into characters, so that the intelligent call can recommend dialogues, whether the intelligent quality inspection is in compliance or not, and the like. The voice recognition model has the effects that a large amount of labeled data of rich scenes are needed for support, data provided by each client terminal is difficult to cover a large amount of service scenes, and the cost of labeling a large amount of data is difficult to bear, so that the accuracy rate of voice recognition is low.

Finally, the user information provided by the agent assistant is not comprehensive enough. Still taking the banking industry as an example, the basic information, financial labels, behavior labels, etc. of the user are needed to comprehensively know the customer, so that the customer service can be completed more efficiently and accurately. Effective and comprehensive user information is difficult to obtain by a single company or organization, the data acquisition cost is high, and common enterprises are difficult to bear.

Based on this, the embodiments of the present invention provide a data processing scheme to solve the above problems.

Fig. 3 is a schematic flowchart of a data processing method according to an embodiment of the present invention, and as shown in fig. 3, the method may include:

and S31, obtaining model parameters respectively sent by a plurality of clients, wherein the model parameters sent by each client are model parameters obtained after the client trains a global model according to local training data, and the local training data are data obtained in the intelligent seat service of the corresponding client.

The execution main body in the embodiment of the invention can be an agent assistant, and the client can be any terminal participating in model training. Each client is provided with respective local training data, and for any client, the local training data is data obtained by performing intelligent seat service on the client, and the local training data corresponding to different clients may be different.

In the process of training the model, firstly, the global model is respectively sent to each client, each client trains the obtained model according to local training data of the client, and the trained model is obtained and sent to the assistant. The model parameters may refer to any parameters for determining the model, and may include, for example, direct parameters in the model, or may include any other parameters for determining direct parameters in the model. The model and the model parameters have corresponding relations, and the model can be correspondingly determined according to the model parameters. The model sending between the client and the agent assistant can be realized by sending model parameters.

In the embodiment of the present invention, the agent assistant may first send the model parameters of the global model to each client, each client updates the local model according to the model parameters, trains the local model by using local training data to obtain the model parameters of the trained local model, and then reports the model parameters of the local model to the agent assistant, and the agent assistant may obtain the corresponding model parameters of the trained local model from each client.

And S32, carrying out aggregation processing on the model parameters sent by the plurality of clients to obtain an updated global model.

After the model parameters respectively sent by the multiple clients are obtained, the model parameters of the multiple clients may be aggregated to obtain the updated global model, and the aggregation mode is not limited, for example, the model parameters of the multiple clients may be aggregated to obtain the model parameters of the updated global model, and then the updated global model is obtained.

In the model training, a plurality of iteration processes may be performed, and the scheme in the embodiment of the present invention may be applied to any one or more iteration processes in the model training, for example, for any one iteration process, the global model may be updated by using the above method, so as to obtain an updated global model.

And S33, processing the input information according to the updated global model to obtain a corresponding auxiliary result, wherein the auxiliary result is used for assisting in determining a response reply corresponding to the input information.

When the user communicates with the seat, the seat assistant may obtain input information, where the input information may include different types, such as voice information when the seat communicates with the user, text information input by the seat, identity information of the obtained user, and the like.

After the input information is acquired, the agent assistant may process the input information through the global model obtained by training to obtain a corresponding assistant result. The type of the auxiliary result may also differ from one type of input information to another. For example, when the input information is the identity information of the user, the auxiliary result may be a representation of the user; when the input information is input text information, the auxiliary result may be a search result corresponding to the text information, and the like.

After the auxiliary result is obtained, the agent may reply for determining a response corresponding to the input information according to the auxiliary result. For example, if the input information is the user's voice information, ask "where do your company have a cable to go off site? After the seat assistant acquires the voice information of the user, voice recognition is carried out on the voice information, so that a search answer of the problem is obtained and displayed to the seat. After seeing the corresponding search answer, the seat can quickly answer the question of the user.

The data processing method provided by the embodiment of the invention comprises the steps of firstly, obtaining model parameters respectively sent by a plurality of clients, wherein the model parameter sent by each client is a model parameter obtained after the client trains a global model according to local training data, and the local training data is data obtained in the intelligent seat service of the corresponding client; then, carrying out aggregation processing on the model parameters sent by the plurality of clients to obtain an updated global model; and processing the input information according to the updated global model to obtain a corresponding auxiliary result, wherein the auxiliary result is used for assisting to determine a response reply corresponding to the input information. In the scheme of the embodiment of the invention, as the global model is obtained by performing federal learning on the respective local training data of the plurality of clients, and the sample data sources are wider and more comprehensive, the common global model obtained by training the sample data provided by the plurality of clients is higher in accuracy and integrity of the model and higher in accuracy of the obtained auxiliary result on the premise of not damaging the privacy of each client, and is more beneficial to response assistance of the assistant to the assistant.

The following describes embodiments of the present invention in detail with reference to the accompanying drawings.

Since the global model in the embodiment of the present invention is obtained by performing federal learning on local training data provided by a plurality of clients, the training of the federal learning model is introduced first.

Fig. 4 is a schematic diagram of a system architecture for federal learning model training provided in an embodiment of the present invention, and as shown in fig. 4, taking a scenario of two data owners a and B as an example, the business systems of a and B respectively own relevant data of respective users, and B also owns tag data to be predicted by model training.

Due to data privacy protection, A and B cannot directly exchange data, so that a federal learning system can be adopted to establish a model. As shown in fig. 4, the system architecture of the federal learning model training mainly includes the following parts:

first, the encrypted samples are aligned.

Since the user groups in the two business systems A and B are not completely overlapped, the shared users of the two business systems A and B are confirmed on the premise that the A and B do not disclose respective data, and the users which are not overlapped with each other are not exposed so as to utilize the characteristics of the users for modeling.

And in the second part, training the encryption model.

After the common user population is determined, the data can be used for training of the federal learning model. In order to ensure the privacy of data in the training process, a third party C is required to assist in encryption training. The training process comprises the following steps:

(1) and the third party C distributes the public key to the A and the B for encrypting the data needing interaction in the training process of the model.

(2) Intermediate results for calculating the gradient are exchanged between a and B in encrypted runs.

(3) A and B are calculated based on the encrypted gradient values respectively, meanwhile, B calculates loss according to the label data of the A and B, and sends the result to a third party C, and the third party C calculates the total gradient value through the summary result and decrypts the total gradient value to obtain the decrypted gradient value.

(4) And the third party C respectively transmits the decrypted gradient values back to A and B, and the A and B respectively update the parameters of the respective models according to the received gradient values.

And repeating the steps until the calculated loss function is converged, and finishing the training process of the whole global model of the federal learning. In the process of user sample alignment and model training, the respective user data of A and B are kept locally, and the privacy and the safety of the data are guaranteed.

Fig. 4 illustrates a basic training process of the federal learning model, which includes horizontal federal learning model training and vertical federal learning model training, which are described below.

Fig. 5a is a schematic diagram of the training of the horizontal federated learning model provided in the embodiment of the present invention, and as shown in fig. 5a, the server and n clients may participate in the federated learning process. Fig. 5b is a schematic diagram of longitudinal federal learning model training provided in the embodiment of the present invention, and as shown in fig. 5b, the server, the client 1, and the client 2 may participate in the federal learning process.

The horizontal federal learning and the vertical federal learning are mainly distinguished for different data sets. When the user features in the data sets provided by the multiple clients overlap more and the user overlap less, the data sets can be divided according to the horizontal direction (namely the user dimension), and data with the same user features but not completely the same users in each client are taken out to train the federated learning model, which is called horizontal federated learning. In the example of fig. 5a, the training process of the horizontal federal learning model is illustrated, which mainly includes 4 steps: (1) each client sends the encrypted gradient value to the server; (2) the server carries out security aggregation; (3) the server sends model updating parameters to each client; (4) and each client updates the model according to the received model updating parameters. The steps are repeatedly executed, and the training of the model can be completed.

When the users of the clients overlap more and the user features overlap less, the data sets of the clients can be divided according to the longitudinal direction (namely feature dimensions), and the data of the clients, which are the same in user and not identical in user features, are extracted for training the federal learning model, which is called longitudinal federal learning. In the example of fig. 5b, the training process of the longitudinal federal learning model is illustrated, which mainly includes 6 steps: (1) each client encrypts the global ID and then transmits the encrypted global ID to the server; (2) the server performs global ID alignment; (3) the server sends training sample IDs to the clients; (4) the encrypted vector conversion (embedding) is sent to a server; (5) the server calculates the encrypted loss and then sends the loss to each client; (6) each client updates the model according to the received losses. The steps are repeatedly executed, and the training of the model can be completed.

In the embodiment of the present invention, the global model may be trained by using the horizontal federal model illustrated in fig. 5a, or may be trained by using the vertical federal model illustrated in fig. 5 b. The plurality of clients perform parameterization processing on respective local training data to obtain corresponding parameter data, and send the respective parameter data to the server. After the server obtains the parameter data respectively sent by the plurality of clients, the server can carry out federal learning training according to the parameter data respectively sent by the plurality of clients to obtain a global model.

After the federal learning training is performed to obtain the updated global model, the input information can be processed according to the updated global model to obtain an intermediate result, and then a corresponding auxiliary result is determined according to the intermediate result.

Since the global model is applied to the agent assistant, the agent assistant may relate to different fields such as classification, speech recognition, semantic understanding, user portrayal, etc., the number of the global model may be multiple, and may include a classification model, a speech recognition model, a semantic understanding model, and a user portrayal model, respectively, for example. Which will be described separately below.

When the global model is a classification model, multiple clients may provide their own local training data for the classification model. For any client, the local training data comprises multiple groups of first sample data, and each group of first sample data comprises first sample information and sample classification labels of the first sample information. The classification system of different clients under their own business system may be different. Taking the banking industry as an example, bank A is an online bank and relates to various online banking services, and bank B relates to both online services and offline network services. If only the data of bank A is used for model training, the trained classification model may not relate to offline system classification. However, if the model training is carried out by combining the data of the bank A and the bank B, the classification model obtained by training can cover the on-line and off-line system classification, so that the classification model is more complete.

The input information can be processed according to the classification model to obtain a classification label corresponding to the input information. And then, according to the classification label corresponding to the input information, determining a classification result corresponding to the classification label in the first database.

Fig. 6 is a first schematic diagram illustrating the assistance result of the seat assistant according to the embodiment of the present invention, as shown in fig. 6, a user 11 and a seat person 12 are in conversation, and a seat assistant 13 is disposed beside the seat person 12. The attendant 12 is a bank employee and is responsible for providing services such as consultation response for the client.

When the user 11 and the attendant 12 are in a call, the attendant 13 can obtain the call between the user 11 and the attendant 12, as the information sent by the user 11 is illustrated in fig. 6, and the user 11 asks "ask you what a good online financial product can be recommended? ", the message is an input message.

The input information is processed through the classification model, and a corresponding classification label can be obtained. For example, according to the "online financing product", the classification label 1 may be "financing" and the classification label 2 may be "online". According to the classification labels 'financing' and 'online', the classification result of the classification label is determined in the first database, as shown by a classification result 60 in fig. 6, the classification result comprises a financing product 1, a financing product 2 and a financing product 3, and the 3 financing products are all online financing products and correspond to the classification labels 'financing' and 'online'. Optionally, the respective characteristics of the 3 financial products can also be illustrated. For example, in fig. 6, financial product 1 is a high-rate and high-risk product, financial product 2 is a medium-rate and medium-risk product, and financial product 3 is a low-rate and low-risk product. The classification result determined according to the classification label 1 and the classification label 2 can be displayed to the attendant 12, and the attendant 12 can specifically introduce the online financial products for the user according to the displayed classification result.

In the embodiment of the invention, the intention classification is improved into the federal intention classification through federal learning. For example, for a single data party, if the services related to the data party are all offline services, there may not be an "online" classification, and the classification model obtained through federal learning has more comprehensive classification labels, so that a more targeted classification result can be obtained.

When the global model is a semantic understanding model, for the semantic understanding model, because multiple different descriptions exist in the same semantic meaning, the intelligent agent service related to a single client often only relates to a few problem sets, and the problem sets are split. The method has the advantages that the local training data provided by the plurality of clients are obtained to carry out model training, so that a more comprehensive problem set can be covered, and the recall rate and the accuracy rate of semantic matching are improved. For any client, the local training data comprises multiple groups of second sample data, and each group of second sample data comprises sample text information and a sample semantic label of the sample text information.

When the input information is text information, the text information can be processed according to the semantic understanding model to obtain a semantic tag corresponding to the text information. And then, determining at least one search answer corresponding to the semantic label in the second database according to the semantic label corresponding to the text information.

Fig. 7 is a schematic diagram illustrating an assistance result of an agent assistant according to an embodiment of the present invention, as shown in fig. 7, a user 11 and an agent person 12 are in conversation, and an agent assistant 13 is disposed beside the agent person 12. The attendant 12 is a bank employee and is responsible for providing services such as consultation response for the client.

When the user 11 and the attendant 12 talk, the user 11 asks "what is the interest of the financial product 3? The attendant 12 may input a question asked by the user as a text, the attendant assistant 13 obtains the text information, and processes the text information through the semantic understanding model to obtain a corresponding semantic tag, which may be, for example, an "interest calculation method for the financial product 3". Then, according to the semantic label, at least one corresponding search answer is determined in the second database. For example, in fig. 7, a search answer 70 is illustrated, and the annual profit rate of the financial product 3 is shown as a% in the search answer 70, which is a general introduction to the interest calculation method for the financial product 3. The search answer 70 also shows specific interest in different amounts of the financial product 3, for example, an interest in one year of saving one thousand is X, an interest in one year of saving ten thousand is Y, and an interest in one year of saving hundred thousand is Z. These several search answers are specific calculations of interest for financial product 3. The agent person 12 can make a specific introduction of interest calculation of the financial product 3 for the user 11 based on the search answer 70.

In the embodiment of the invention, the semantic matching is improved into the federated semantic matching through the federated learning. For example, for a single data party, the data party refers to the interest inquiry method of "how much interest is for a financial product", but there may be other inquiry methods, such as "how much interest is for ten thousand money for one year". And the semantic understanding model obtained through the federal learning can contain more problem sets, so that the corresponding search answer can be determined according to the obtained semantic tag.

When the global model is a speech recognition model, for the speech recognition model, the speech recognition model needs large labeling data under rich scenes for support, and intelligent agent services related to a single client are difficult to cover a large number of service scenes. The model training is carried out by acquiring the local training data provided by the plurality of clients jointly, so that the training data can cover more and more comprehensive scenes, the obtained voice recognition model is more perfect and comprehensive, and the voice recognition effect is better. For any client, the local training data comprises multiple groups of third sample data, and each group of third sample data comprises sample voice information and a sample voice recognition result of the sample voice information.

When the input information is voice information, the voice information can be processed according to the voice recognition model, and a voice recognition result corresponding to the voice information is obtained. And then, determining at least one response text corresponding to the voice recognition result in a third database according to the voice recognition result corresponding to the voice information.

Fig. 8 is a third schematic diagram illustrating the assistance result of the seat assistant according to the embodiment of the present invention, as shown in fig. 8, the user 11 and the seat person 12 are in conversation, and a seat assistant 13 is disposed beside the seat person 12. The attendant 12 is a bank employee and is responsible for providing services such as consultation response for the client.

When the user 11 and the attendant 12 talk, the user 11 asks "ask you what better online financing product can be recommended? "the input information sent by the user is a voice message, and the agent assistant 13 can obtain the voice message and process the voice message through the voice recognition model to obtain a corresponding voice recognition result, so as to know that the user 11 wants to know about the online financial product. Then, based on the speech recognition result, a corresponding at least one answer text may be determined in a third database. For example, in fig. 8, a reply text 80 is shown, which includes 4 alternative reply schemes. Response 1: good I introduce a few products for your consideration. Response 2: somehow, do you not know what financial product you want on this side? Response 3: do you have a requirement on these aspects of profitability? Response 4: this side recommends several products for you based on your historical consumption records.

The several alternative answer texts may be presented to the agent person 12 for the agent person 12 to select any one of them to quickly answer the user 11.

When the global model is the user portrait model, the user portrait model is trained through local training data provided by a plurality of clients, more user label information under each user identity information can be obtained, and accordingly more accurate user labels can be constructed. For any client, the local training data comprises multiple groups of fourth sample data, and each group of fourth sample data comprises sample user information and a sample user identifier of the sample user information.

When the input information is user information, the user information can be processed according to the user portrait model to obtain a user identifier corresponding to the user information. And then, according to the user identification corresponding to the user information, determining the user label corresponding to the user identification in a fourth database.

Fig. 9 is a fourth schematic diagram illustrating the assistance result of the seat assistant according to the embodiment of the present invention, as shown in fig. 9, the user 11 and the seat person 12 are in conversation, and a seat assistant 13 is disposed beside the seat person 12. The attendant 12 is a bank employee and is responsible for providing services such as consultation response for the client.

When the user 11 and the attendant 12 talk, the attendant assistant 13 acquires the telephone number of the user 11. Typically, the phone number of the user is bound with the relevant information of the user, so that the phone number is a piece of user information that can be the user 11.

After the assistant 13 obtains the phone number of the user 11, the phone number of the user 11 may be input into the global model, and the user identification of the user 11 output by the global model is obtained, for example, the name of the user 11 illustrated in the page 90 in fig. 9 is Li Ming, the phone number is 13XXXXXXXXX, and the address is Beijing, etc. Then, based on the above information of the user 11, a user tag of the user may be obtained, for example, in the page 90, the historical consumption records of the user 11 in the past year are 13 times, including record 1, record 2, record 3, and so on, and the user tag may be a low-risk consumer based on the historical consumption records of the user 11 in the past year in combination with other information, and so on.

In the example of FIG. 9, the user representation model is trained from local training data provided by a plurality of clients. Taking the banking industry as an example, assume that the user 11 holds a bank card of bank a and has consumed it over the past year. The seat staff 12 is a staff of bank B and cannot know the consumption data of the user 11 on the bank card a. If the global model is trained only by using sample data provided by bank B, the user portrait model cannot acquire the consumption record of the user 11 in bank A. Because the consumption record in bank A cannot be obtained, the label of the user cannot be jointly determined by combining the consumption record of bank A.

And if the user portrait model is obtained by training sample data provided by a plurality of clients, the client can parameterize various information of the same user identity information by using a client mechanism by means of a federal learning mechanism and transmit the information back to the server, and the server performs joint training by combining the parameterized information provided by the clients to obtain the user portrait model, so that the user portrait model can obtain a more perfect user label according to the user information.

The assistant can provide services such as user portrayal, intelligent recommendation, answer search and the like, and can also provide intelligent quality inspection services. The agent assistant can perform real-time quality inspection aiming at the dialog text of the quality inspection of the user and the agent according to the quality inspection item requirement for realizing the configuration, feed back the quality inspection result to the agent, prompt the attention items in the dialog of the agent in time and supervise and standardize the behavior of the agent. When an agent is found to have an inappropriate utterance, e.g., an abuser, terror, etc., then a warning may be controlled, e.g., "please note your utterance! "and the like.

Fig. 10 is a schematic structural diagram of a data processing apparatus according to an embodiment of the present invention, and as shown in fig. 10, the apparatus includes:

the obtaining module 101 is configured to obtain model parameters sent by multiple clients, where the model parameter sent by each client is a model parameter obtained after the client trains a global model according to local training data, and the local training data is data obtained in an intelligent seat service performed by the corresponding client;

the training module 102 is configured to perform aggregation processing on the model parameters sent by the multiple clients to obtain an updated global model;

and the processing module 103 is configured to process the input information according to the updated global model to obtain a corresponding auxiliary result, where the auxiliary result is used to assist in determining a response reply corresponding to the input information.

In a possible implementation manner, the processing module 103 is specifically configured to:

In one possible implementation, the global model is a classification model, the local training data includes a plurality of groups of first sample data, each group of first sample data includes first sample information and a sample classification label of the first sample information; the processing module 103 is specifically configured to:

In a possible implementation manner, the input information is text information, the global model is a semantic understanding model, the local training data includes a plurality of sets of second sample data, and each set of second sample data includes sample text information and a sample semantic tag of the sample text information; the processing module 103 is specifically configured to:

In a possible implementation manner, the input information is speech information, the global model is a speech recognition model, the local training data includes a plurality of sets of third sample data, and each set of third sample data includes sample speech information and a sample speech recognition result of the sample speech information; the processing module 103 is specifically configured to:

In a possible implementation manner, the input information is user information, the global model is a user portrait model, the local training data includes a plurality of sets of fourth sample data, and each set of fourth sample data includes sample user information and a sample user identifier of the sample user information; the processing module 103 is specifically configured to:

The data processing apparatus provided in any of the foregoing embodiments is configured to execute the technical solution of any of the foregoing method embodiments, and the implementation principle and the technical effect are similar, which are not described herein again.

Fig. 11 is a schematic structural diagram of a data processing device according to an embodiment of the present invention. As shown in fig. 11, the apparatus may include: a memory 111, a processor 112 and a data processing program stored on the memory 111 and operable on the processor 112, the data processing program, when executed by the processor 112, implementing the steps of the data processing method according to any of the preceding embodiments.

Alternatively, the memory 111 may be separate or integrated with the processor 112.

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

An embodiment of the present invention further provides a computer-readable storage medium, where a data processing program is stored on the computer-readable storage medium, and when the data processing program is executed by a processor, the data processing program implements the steps of the data processing method according to any of the foregoing embodiments.

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

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

It should be understood that the Processor may be a Central Processing Unit (CPU), other general purpose Processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), etc. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of a method disclosed in connection with the present invention may be embodied directly in a hardware processor, or in a combination of the hardware and software modules within the processor.

The memory may comprise a high-speed RAM memory, and may further comprise a non-volatile storage NVM, such as at least one disk memory, and may also be a usb disk, a removable hard disk, a read-only memory, a magnetic or optical disk, etc.

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

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

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

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

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

The above description is only a preferred embodiment of the present invention, and not intended to limit the scope of the present invention, and all modifications of equivalent structures and equivalent processes, which are made by using the contents of the present specification and the accompanying drawings, or directly or indirectly applied to other related technical fields, are included in the scope of the present invention.

Claims

1. A method of data processing, the method comprising:

2. The method according to claim 1, wherein processing the input information according to the updated global model to obtain a corresponding auxiliary result comprises:

3. The method of claim 2, wherein the global model is a classification model, the local training data comprises a plurality of sets of first sample data, each set of first sample data comprising first sample information and sample classification labels for the first sample information; the processing the input information according to the updated global model to obtain an intermediate result includes:

4. The method of claim 3, wherein said determining the corresponding auxiliary result from the intermediate result comprises:

5. The method of claim 2, wherein the input information is text information, the global model is a semantic understanding model, the local training data comprises a plurality of sets of second sample data, each set of second sample data comprises sample text information and sample semantic tags of the sample text information; the processing the input information according to the updated global model to obtain an intermediate result includes:

6. The method of claim 5, wherein said determining the corresponding auxiliary result from the intermediate result comprises:

7. The method of claim 2, wherein the input information is speech information, the global model is a speech recognition model, the local training data comprises a plurality of sets of third sample data, each set of third sample data comprises sample speech information and a sample speech recognition result of the sample speech information; the processing the input information according to the updated global model to obtain an intermediate result includes:

8. The method of claim 7, wherein determining the corresponding auxiliary result from the intermediate result comprises:

9. The method of claim 2, wherein the input information is user information, the global model is a user profile model, the local training data comprises a plurality of sets of fourth sample data, each set of fourth sample data comprises sample user information and a sample user identification of the sample user information; the processing the input information according to the updated global model to obtain an intermediate result includes:

10. The method of claim 9, wherein said determining the corresponding auxiliary result from the intermediate result comprises:

11. A data processing apparatus, characterized in that the apparatus comprises:

12. A data processing apparatus, characterized in that the data processing apparatus comprises: memory, processor and data processing program stored on the memory and executable on the processor, which when executed by the processor implements the steps of the data processing method according to any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that a data processing program is stored thereon, which when executed by a processor implements the steps of the data processing method according to any one of claims 1 to 10.

14. A computer program product, characterized in that it comprises a computer program which, when being executed by a processor, carries out the steps of the data processing method according to any one of claims 1 to 10.