CN118013414A - AI-based loan-assisting business outbound data inspection method, electronic equipment and program product - Google Patents

Abstract

The invention discloses a method for inspecting outbound data of a loan-aid business. The construction method of the inspection model comprises the steps of collecting outbound data of an outbound platform through construction kfk service; cleaning and preprocessing the collected outbound data; and taking the cleaned and preprocessed outbound data as training samples of the inspection model, and extracting features of the training samples. The feature extraction comprises keyword extraction of outbound data by adopting a word bag model, and converting keywords into labels. The inspection model is a transform-based deep learning model.

Description

AI-based loan-assisting business outbound data inspection method, electronic equipment and program product

Technical Field

The invention belongs to the technical field of digital finance, and particularly relates to an AI-based loan-aid business outbound data inspection method, electronic equipment and a program product.

Background

With the continuous development of artificial intelligence technology, intelligent outbound calls have been developed. The intelligent outbound technology belongs to the fields of Natural Language Processing (NLP) and Artificial Intelligence (AI), and the basic principle is to realize natural communication between a system and a person by utilizing a computer technology and a voice recognition technology. The intelligent outbound technology can be applied to multiple fields of customer service, sales, market research and the like, so that the efficiency and the service quality of enterprises are greatly improved, the operation cost is reduced, and the marketing income is improved.

Disclosure of Invention

According to one embodiment of the disclosure, an intelligent outbound data AI (analog) inspection method for a loan-aiding scene based on deep learning is provided, and the inspection is performed on the outbound data of a loan-aiding business through a trained inspection model. The construction method of the inspection model comprises the following steps of,

Collecting outbound data of an outbound platform through setting up kfk services;

cleaning and preprocessing the collected outbound data;

And taking the cleaned and preprocessed outbound data as training samples of the inspection model, and extracting features of the training samples.

The feature extraction comprises keyword extraction of outbound data by adopting a word bag model, and keyword conversion is carried out into labels.

The inspection model is a transform-based deep learning model.

Drawings

The above, as well as additional purposes, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description when read in conjunction with the accompanying drawings. Several embodiments of the present invention are illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings and in which:

fig. 1 is a flow chart of an outbound data inspection method according to one embodiment of the invention.

Detailed Description

As the outbound traffic increases, the outbound data exhibits an exponential rise, making quality management of the outbound traffic data increasingly important. However, at present, the data inspection has many defects and shortcomings, including:

(1) With the increase of the data volume, the workload of manual verification error correction feedback to the intention model in the outbound system is increased, so that the outbound data of the current day cannot be effectively and timely verified and detected, and the follow-up intention model optimization and update are influenced. Meanwhile, the manual data checking has a difference in data checking accuracy due to the fact that personal business knowledge is mastered differently;

(2) In the auxiliary loan service scene, the complaint condition exists, the contents of each call need to be analyzed and generalized, whether the customer has a tendency of complaint in the call process or not is checked, and if the customer relies on manual analysis and check on tens of millions of calls every day, the analysis and the review are obviously not completed.

The auxiliary loan service is called as a loan assisting service for short. The loan-aid business refers to a financial institution that provides financing services to customers in need thereof by cooperating with a third party platform. The business process is mainly divided into three stages of pre-loan, mid-loan and post-loan, and the marketing channel is divided into an online channel and an offline channel. On-line is mainly done by outbound telephony. The conversational experience of the customer is particularly important during the outbound business. Here, a good intelligent outbound system can accurately capture the intention of the customer and respond, answer or inform the customer of the relevant information quickly. This is particularly important in two links: one is the accuracy of an intention recognition model of the outbound system, and the other is to listen to a telephone recording after the outbound call is ended, carry out data inspection, check whether to solve the problem of the client, avoid complaints and improve the satisfaction of the client.

In the aspect of improving the intention model and accurately identifying the intention of a customer, the conventional method is to manually perform inspection and screening. Here, the operator who needs to have deep knowledge about the scene of the loan-aid business works, and first, the staff business training cost is high, and the novice needs to train for three months to work on duty. And the service content of each loan-aid service line has dozens of service intentions and service problems, so that the complexity is high, and time and labor are wasted when professional staff perform comparison and check. And because the business levels of different business personnel are uneven, the checking and checking out operation can have different conditions, and after the outbound data is checked and checked for a long time, the working efficiency of the staff can be obviously reduced, and errors are easy to check. Finally, the dialects are marked by marking personnel and then fed back to the intention model training to improve the recognition effect. The labor cost consumed in the process is high, and the time is very long when the number of outgoing calls is large.

In terms of improving customer satisfaction and solving customer problems to avoid complaints, the conventional method is performed manually or through simple rule matching, so that time and labor are wasted, and the problem of increased complaints is easily generated. If the problem of the customer is not solved by the one-way intelligent outbound call, the customer is not satisfied with the call by checking and checking manually or by simple rule matching, or by checking the call for a long time, the customer considers that the problem of the customer is not solved or is played by a robot, and complaints are caused.

The purpose of the present disclosure is to provide an intelligent outbound data inspection method based on deep learning, which inspects outbound data in an automated manner, and improves the accuracy and reliability of outbound data, thereby improving the efficiency and quality of the lending-aiding business.

According to one or more embodiments, an outbound data inspection method includes the steps of:

And (5) data acquisition.

The outbound platform has tens of millions of outbound data output every day, and outbound data collection is carried out by constructing kfk services.

Data cleaning and preprocessing.

In an actual outbound scenario, the collected data may include a lot of noise and invalid information, thus requiring data cleaning and preprocessing first. The collected data is cleaned, including operations such as deduplication, formatting, standardization and the like, so as to ensure the accuracy and consistency of the data.

Feature extraction translates to labels.

The complaint module in AI inspection requires feature extraction for the collected data to be modeled and analyzed.

The method adopts the word bag model to carry out keywords on outbound data, and the keywords are converted into labels after being extracted, wherein the labels are as follows:

data_sum= { ' overdrive behavior ':5, ' how much work: 73, ' special case consultation ': the automated is automated needs as automated needs as needs, is, as, of, what is the equity black card ' 259, ' password question ' 4, ' challenge overdue days ' 6, ' payback limited ' 15}

Data_error= { 'challenge robot': 142, 'consulting identity': 136, 'voice assistant': the automated electric back-up of 125, the manual electric back-up of 96, the automated electric back-up of 54, the automated electric back-up of 43, the automated electric back-up of 39, the automated electric back-up of 33, the latest electric back-up of 30, the automated electric back-up of 21, the automated electric back-up of 20, the automated electric back-up of 1, the automated electric back-up of 19, the automated electric back-up of 14, the automated electric back-up of 11, the automated back-up of 10, the automated electric back-up of 9, the automated back-up of 8, the automated electric back-up of 8, the automated back-up of 6, the automated back-up of 4, the automated back-up of 1, the automated back-up to 6, the automated back-up to the automated back-up of the automated device, the automated back-up to the automated device of the automated device

The tag weight calculation score is performed by the following formula:

scores＝data_error/data_sum

And (5) establishing a model.

Because of the data characteristics of the outbound dialogue, in the whole conversation process, some clients have incomplete speaking, repeated speaking and intermittent response, so that information is incomplete and important information is not highlighted, the invention adopts a method for constructing a deep learning model based on a transducer, and the transducer is a deep learning model based on a self-attention mechanism, so that the mechanism can reduce information loss: the self-attention mechanism can help the neural network to pay attention to important information in the input sequence, so that information loss is reduced, and accuracy of a model is improved.

Some clients have complete speaking and speak several sentences, which leads to overlong text input to the model, and the problem of inaccurate recognition exists in the common model, but the deep learning model is constructed based on a transducer, so that the generalization capability of the model can be improved due to the self-attention mechanism: the self-attention mechanism can help the neural network learn the long-range dependency relationship in the input sequence, so that the generalization capability of the model is improved, and the model can better perform when the input sequences with different lengths are processed.

First, the model structure is a Embedding level, mapping each word into a vector representation. The vector sequence then passes through a Transformer Encoder layer, where each Encoder layer is composed of multiple attention mechanisms and feed forward neural networks for extracting important features in the sequence. Finally, after passing through a plurality of Encoder layers, the vector sequence is summarized into a fixed-length vector, and the vector passes through a full-connection layer, and finally the classification result of the sequence is output. Specifically, each Encoder layer contains the following steps: multi-head attention mechanism: the input sequence is mapped into a plurality of queries, keys and value vectors, respectively, and then the attention distribution is calculated for weighted summation of each value vector to obtain an output vector. Residual connection: and adding the input vector and the output vector of the multi-head attention mechanism to obtain a residual vector. Layer normalization: and normalizing the residual vector and accelerating convergence. Feedforward neural network: and performing full-connection operation on the normalized residual vector to obtain a new vector. Residual connection again: and adding the output vector of the feedforward neural network and the normalized residual vector to obtain a new residual vector. And (5) carrying out layer normalization: the new residual vector is normalized. By stacking a plurality of Encoder layers, the high-level features of the input sequence can be extracted step by step, thereby realizing the target service.

And (5) model training.

And marking data in a data collection stage, namely complaint data and non-complaint data respectively, and putting the data into the constructed complaint model for training.

The correction model is characterized in that the training corpus of the correction model is different from the complaint model, the training corpus is directly expressed by using the corpus of the intention model of the intelligent outbound, a model is constructed based on a deep learning transducer structure to carry out embedding, and vector space clustering is carried out through faiss to carry out similarity semantic search.

Model testing and evaluation.

After the model training is completed, it needs to be tested and evaluated.

Complaint model the embodiment of the disclosure predicts by using a ten-day outbound data input model without training, compares the predicted result of the model with the real condition of the data, and evaluates the model for professional operators.

According to the embodiment of the disclosure, seven different intention categories are predicted by inputting a round of conversation of a customer into the model for prediction, the seven intention corpuses are compared with the intention categories recognized by the outbound intention model, if the intention categories recognized by the intention model of the outbound system are not in the seven intention categories, the round of conversation of the customer is data for recognizing errors, the intention categories need to be re-labeled by the customer and a professional operator, and then the intention categories need to be fed back to the intention model for retraining, so that the intention model is optimized, and the outbound effect is better improved.

And (3) data feedback: and feeding back the inspection result to the data source so as to carry out corresponding correction and improvement on the data source.

Deployment and optimization.

Finally, after the model test and evaluation is completed, it can be deployed and optimized. Models can be embedded in practical applications and continually optimized and refined, for example, using reinforcement learning, migration learning, etc. techniques to further enhance their performance and effectiveness.

Through the technical scheme, the intelligent outbound data AI inspection of the lending scene can realize comprehensive monitoring and management of data quality, and the accuracy and consistency of the data are improved, so that the efficiency and accuracy of the lending service are improved.

In view of the above, the technical effects of the present disclosure include,

According to the complaint model in the embodiment of the disclosure, whether the customer complains or not is predicted by converting emotion of the customer, speech semantics of the customer and extracted features into dimensions such as labels, and if the customer with complaint tendency is identified, the first time gives an electric representation, so that complaint cases are avoided. The complaint model has the advantages that 1. The data analysis and check of each call by operators are avoided, and the labor cost is saved;

2. the number of customer complaints per day is reduced, the customer satisfaction is improved, the outbound effect is improved, the number of complaints is reduced, the corresponding operation cost is reduced, and the cost is saved.

According to the error correction model, similarity intention of top7 is found through similarity recognition of each round of call operation of a client, the similarity intention is compared with the recognition result of the intention recognition model of intelligent outbound, and inconsistent comparison is selected and submitted to operators for verification. The error correction model has the beneficial effects that 1. The working efficiency of operators is improved, and each call is not required to be verified;

2. by correcting and feeding back the badcase data identified by the original intelligent outbound intention identification model, the model identification accuracy is improved, the customer answering experience in outbound is optimized, and further the business performance improvement is promoted.

It should be understood that, in the embodiment of the present invention, the term "and/or" is merely an association relationship describing the association object, which means that three relationships may exist. For example, a and/or B may represent: a exists alone, A and B exist together, and B exists alone. In addition, the character "/" herein generally indicates that the front and rear associated objects are an "or" relationship.

The integrated units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention is essentially or a part contributing to the prior art, or all or part of the technical solution may be embodied in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a usb disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

While the invention has been described with reference to certain preferred embodiments, it will be understood by those skilled in the art that various changes and substitutions of equivalents may be made and equivalents will be apparent to those skilled in the art without departing from the scope of the invention. Therefore, the protection scope of the invention is subject to the protection scope of the claims.

Claims (10)

1. A method for inspecting the outbound data of a loan-aiding service is characterized in that the outbound data of the loan-aiding service is inspected through a trained inspection model.

2. The method according to claim 1, wherein the method for constructing the inspection model includes collecting outbound data of an outbound platform by constructing kfk services;

cleaning and preprocessing the collected outbound data;

And taking the cleaned and preprocessed outbound data as training samples of the inspection model, and extracting features of the training samples.

3. The method of claim 2, wherein the feature extraction comprises keyword extraction of outbound data using a word bag model and converting keywords to tags.

4. The method of claim 2, wherein the inspection model is a transform-based deep learning model.

5. The method of claim 4, wherein the inspection model comprises an intent model.

6. The method according to claim 2, wherein a complaint sub-model is provided in the inspection model, complaints or non-complaints are marked in training samples, and the complaint sub-model is trained with the marked or complaint or non-complaint samples.

7. The method of claim 5, wherein an error correction sub-model is provided in the inspection model, the error correction sub-model being trained based on samples of the intent sub-model.

8. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor runs the computer program to implement the method of any one of claims 1 to 7.

9. A storage medium having stored thereon a computer program which, when executed by a processor, implements the method of any of claims 1 to 7.

10. A computer program product comprising a computer program, characterized in that the computer program is executed by a processor to implement the method of any one of claims 1 to 8.