CN112101423A - Multi-model fused FAQ matching method and device - Google Patents

Abstract

The invention is applicable to the technical field of natural language processing, and provides a method and device for FAQ matching of multi-model fusion. By sequentially obtaining the training text set of the problem to be processed, combined with Bert-encoder+DBScan clustering to assist in extracting and summarizing financial management Educating knowledge points to build a financial education FAQ, manually annotating a small number of similar questions, then generating a large number of similar questions based on a small number of similar questions that have been labeled, and doing manual review, constructing a sentence pair matching dataset, using unsupervised models and supervised depth The learning model trains the sentence pair matching model. After the final sentence pair matching model training is completed, the question input by the user is received, and the question that best matches the input question and its corresponding answer output are identified. Train the model, extract text, match standard questions, and reply with corresponding answers to solve the problem of cumbersome viewing by FAQ users and low efficiency of manual customer service.

Description

A method and device for FAQ matching of multi-model fusion

technical field

The invention belongs to the technical field of natural language processing, and in particular relates to a method and device for FAQ matching of multi-model fusion.

Background technique

There are not many cases in the field of artificial intelligence in the financial education industry, especially the private industry corpus on the FAQ, which is inconvenient to open source, which limits the development.

In recent years, the field of NLP has developed rapidly, but there are not many that can be implemented in financial education and achieve good results. The most cutting-edge FAQ sentence pair matching algorithm is not well implemented in the field of financial education.

Intelligent FAQ is difficult to achieve better results when there are many knowledge points and the semantics of the questions are very similar.

SUMMARY OF THE INVENTION

The present invention provides a method and device for FAQ matching of multi-model fusion, aiming at solving the problems existing in the prior art.

The present invention is achieved in this way, a method and device for FAQ matching of multi-model fusion, comprising the following steps:

S1. Obtain the training text set of the problem to be processed, and combine Bert-encoder+DBScan clustering to assist in extracting and summarizing financial education knowledge points, thereby constructing a financial education FAQ, and manually marking a small number of similar problems;

S2. Use the similar problem generation module to generate a large number of similar problems according to a small number of similar problems that have been marked, and do manual review to construct a sentence pair matching data set;

S3. Build a pre-training model dataset;

S4. Use an unsupervised model and a supervised deep learning model to train a sentence pair matching model;

S5. After the training of the sentence pair matching model is completed, the question input by the user is received, and the text of the input question is matched to the sentence pair matching model to identify the question that best matches the input question and its corresponding answer output, and reply to the user.

Preferably, the training text set includes texts with a length limit of 3-50, and texts with expressions, numbers and emails deleted.

Preferably, the unsupervised model includes a WMD model and a SIF model;

The supervised model includes a bert model, an albert model, and a roberta model.

The present invention also provides a multi-model fusion FAQ matching device, comprising:

Financial education corpus database, which is used to store pre-input FAQ corpus data and generate training text sets;

Manual labeling module, which is used for operators to manually label a small number of similar problems in the training text set;

A similar problem generation module, which includes a similar problem generation model, and the similar problem generation model is used to generate a large number of similar problems according to a small number of similar problems that have been marked, and perform manual review to construct a sentence pair matching data set;

The NLU module is used to train a sentence pair matching model, and use the trained model to match the questions input by the user, find the most matching question, output the most matching question and its corresponding answer, and reply to the user.

Preferably, the training text set includes texts with a length limit of 3-50, and texts with expressions, numbers and emails deleted.

Preferably, the NLU module includes an unsupervised model and a supervised model;

Described unsupervised model includes WMD model, SIF model;

The supervised model includes a bert model, an albert model, and a roberta model.

Compared with the prior art, the beneficial effects of the present invention are: a method and device for FAQ matching of multi-model fusion of the present invention, by sequentially obtaining the training text set of the problem to be processed, combined with Bert-encoder+DBScan clustering assistance Extract and summarize financial education knowledge points to build a financial education FAQ, manually mark a small number of similar questions, and then generate a large number of similar questions based on a small number of similar questions that have been marked, and do manual review to construct a sentence pair matching dataset, using unsupervised The model and the supervised deep learning model train the sentence pair matching model. After the final sentence pair matching model training is completed, the question input by the user is received, and the question that best matches the input question and its corresponding answer output are identified. The present invention adopts multiple models. Fusion, training or pre-training models, extracting text, matching standard questions, and replying to the corresponding answers to solve the problem of cumbersome FAQ users and low efficiency of manual customer service.

Description of drawings

FIG. 1 is a schematic diagram of an overall system of a multi-model fusion FAQ matching device of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

Referring to Fig. 1, the present invention provides a technical solution: a method and device for FAQ matching of multi-model fusion, and the method for FAQ matching of multi-model fusion comprises the following steps:

S1. Obtain the training text set of the problem to be processed, and combine Bert-encoder+DBScan clustering to assist in extracting and summarizing financial education knowledge points, thereby constructing a financial education FAQ, and manually annotating a small number of similar problems. Among them, the training text set includes texts with a text length limit of 3-50, and texts with expressions, numbers, and emails deleted.

In this embodiment, the question is: what is the difference between A shares and B shares?

The similar questions for artificial increase are: ① How to distinguish between A shares and B shares? ②A shares are better or B shares are better?

S2. Use the similar problem generation module to generate a large number of similar problems according to a small number of similar problems that have been marked, and do manual review to construct a sentence pair matching data set.

In this embodiment, the sentence pair matching dataset is shown in Table 1 below:

Table 1

standard question similar problem What is the difference between A shares and B shares? How to distinguish between A shares and B shares? I want to know if A shares are the same as B shares? Are A shares the same as B shares? Hello, may I ask what concept stocks are? What is the difference between A shares and B shares? What is concept stock? What is concept stock? What is the meaning of concept stocks? What is concept stock?

S3. Build a pre-training model data set according to the sentence pair matching data set.

S4. Use an unsupervised model and a supervised deep learning model to train a sentence pair matching model. Unsupervised models include WMD models and SIF models. Supervised models include bert model, albert model, and roberta model. Further predict the probability according to the model, and use the linear regression + XGBOOST solution online, in which the linear regression can train and update the model in real time, accounting for 40%, and XGBOOST uses the pre-trained model, accounting for 60%. In this way, 60% of the XGBOOST guarantee The stability of the model is guaranteed, and the 40% linear regression ensures the flexibility of the model.

S5. After the training of the sentence pair matching model is completed, the question input by the user is received, and the text of the input question is matched to the sentence pair matching model to identify the question that best matches the input question and its corresponding answer output, and reply to the user.

A multi-model fusion FAQ matching device of the present invention includes a manual labeling module, a similar problem producing module, and an NLU module.

The financial education corpus database is used to store pre-input FAQ corpus data and generate training text sets. The training text set includes texts with a text length limit of 3-50, removing expressions, numbers and emails.

The manual labeling module is used for operators to manually label a small number of similar problems in the training text set;

The similar question generation module includes a similar question generation model, and the similar question generation model is used to generate a large number of similar questions according to a small number of similar questions that have been marked, and conduct manual review to construct a sentence pair matching data set.

The NLU module is used to train a sentence pair matching model, and use the trained model to match the questions input by the user, find the most matching question, output the most matching question and its corresponding answer, and reply to the user. NLU modules include unsupervised models and supervised models. Unsupervised models include WMD models and SIF models. Supervised models include bert, albert, and roberta.

The above modules are all deployed in an online environment on two RTX60024G GPU servers. Online environment, optimize service performance to solve high concurrency problems, and control the response speed within 300ms. The optimization process includes: parallel computing of the bert preprocessing process, hot loading of the trained model, and parallel computing of multiple models.

A method and device for FAQ matching of multi-model fusion of the present invention, by sequentially obtaining the training text set of the problem to be processed, combined with Bert-encoder+DBScan clustering to assist in extracting and summarizing financial education knowledge points, thereby constructing a financial education FAQ , manually mark a small number of similar problems, then generate a large number of similar problems according to the small number of similar problems that have been marked, and do manual review, build a sentence pair matching dataset, use an unsupervised model and a supervised deep learning model to train the sentence pair matching model, and finally After the training of the sentence pair matching model is completed, the question input by the user is received, and the question that best matches the input question and its corresponding answer output are identified. , and reply with the corresponding answer to solve the problem that FAQ users are cumbersome to view and the efficiency of manual customer service is low.

The above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims (6)

1. A multi-model fused FAQ matching method is characterized in that: the method comprises the following steps:

s1, obtaining a training text set of the problems to be processed, extracting and concluding the finance education knowledge points in combination with Bert-encoder + DBSCAN clustering assistance, thereby constructing a finance education FAQ and manually labeling a small number of similar problems;

s2, generating a large number of similar problems by using a similar problem generation module according to a small number of labeled similar problems, manually checking and constructing a sentence pair matching data set;

s3, constructing a pre-training model data set;

s4, training sentence pair matching models by using an unsupervised model and a supervised deep learning model;

and S5, after the sentence pair matching model is trained, receiving the questions input by the user, inputting the texts of the questions into the sentence pair matching model, identifying the questions most matched with the input questions and corresponding answers, outputting the questions and the corresponding answers, and replying the answers to the user.

2. The method of multi-model fused FAQ matching as claimed in claim 1, wherein: the training text set includes text length limits between 3-50, deleted expressions, numbers, and text of the email.

3. The method of multi-model fused FAQ matching as claimed in claim 1, wherein: the unsupervised model comprises a WMD model and an SIF model;

the supervised model comprises a bert model, an albert model and a roberta model.

4. An apparatus for multi-model fused FAQ matching, comprising: the method comprises the following steps:

the financial education corpus database is used for storing pre-input FAQ corpus data and generating a training text set;

the manual labeling module is used for manually labeling a small number of similar problems in the training text set by an operator;

the similar problem generation module is used for generating a large number of similar problems according to a small number of labeled similar problems, carrying out manual review and constructing a sentence pair matching data set;

and the NLU module is used for training the sentence pair matching model, matching the questions input by the user by using the trained model, finding out the best matching question, outputting the best matching question and the corresponding answer thereof, and replying the best matching question and the corresponding answer to the user.

5. The apparatus for multiple model fused FAQ matching according to claim 4, wherein: the training text set includes text length limits between 3-50, deleted expressions, numbers, and text of the email.

6. The apparatus for multiple model fused FAQ matching according to claim 4, wherein: the NLU module comprises an unsupervised model and a supervised model;

the unsupervised model comprises a WMD model and an SIF model;

the supervised model comprises a bert model, an albert model and a roberta model.

Images