Uyghur language consistency judgment method and system in credit review

Abstract

The invention discloses a Uyghur language consistency determination method in credit review, and the method comprises the steps: (1) outputting a current question to a credit review user through a question and answer robot, collecting a Uyghur language voice answer of the credit review user, and converting the Uyghur language voice answer into a Uyghur language text; (2) obtaining a Chinese text of a corresponding answer of the current question in application information submitted by the credit review user; and (3) inputting the Uygur text corresponding to the current question and the Chinese text into the trained judgment model, judging whether the current Uygur text output by the model is consistent with the semantics of the Chinese text, and outputting a judgment result. Based on the judgment model formed by the deep learning technology, the input Uyghur language text and Chinese text are directly subjected to semantic consistency recognition, the Uyghur language does not need to be translated into Chinese and then subjected to consistency judgment, the consistency recognition speed and recognition precision are greatly improved, and online Uyghur language and Chinese semantic consistency recognition can be achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of letter review, and more specifically, the present invention relates to a method and system for determining Uyghur language consistency in letter review. Background Art

[0002] During the credit approval process, the credit reviewer needs to conduct an electronic review with the credit review user over the phone. During the electronic review process, the credit reviewer needs to ask questions to the user, verify the customer's application materials based on the user's answers, and conduct the next round of questions based on the customer's answers.

[0003] In-depth questions in the credit review dialogue are crucial to understanding the client's qualifications. The current credit market customer base is becoming increasingly diverse, including customers from different industries, professions and languages. When the credit review user uses Uyghur to communicate with the credit review officer, the Uyghur voice response collected from the credit review user is first converted into Uyghur text, and then the Uyghur text is translated into Chinese text. Based on existing methods, it is determined whether the credit review user's answer is consistent with the relevant Chinese text in the submitted application information.

[0004] It is necessary to convert Uyghur into Chinese before making a consistency judgment. Converting Uyghur into Chinese requires the use of a translation model. Training a Uyghur to Chinese translation model also requires annotating a large amount of text data, and the cost of this annotation is also very high. In addition, the translation model itself has an upper limit on accuracy, and making similarity judgments after translation will result in the problem of error accumulation. Summary of the invention

[0005] The present invention provides a method for determining consistency of Uyghur language in letter review, aiming to improve the above-mentioned problem.

[0006] The present invention is implemented as follows: a method for determining the consistency of Uyghur language in letter review, the method comprising the following steps:

[0007] (1) The question-answering robot outputs the current question to the credit review user, collects the Uyghur voice answer from the credit review user, and converts the Uyghur voice answer content into Uyghur text;

[0008] (2) Obtain the Chinese text of the answer to the current question in the application information submitted by the credit review user;

[0009] (3) The Uyghur text and the Chinese text corresponding to the current question are input into the trained judgment model. The judgment model outputs whether the current Uyghur text is semantically consistent with the Chinese text and outputs a judgment result.

[0010] Furthermore, the determination model includes:

[0011] The first CINO model and the second CINO model;

[0012] A Concat layer connected to the first CINO model and the second CINO model;

[0013] Fusion layer connected to Concat layer;

[0014] The input and output of the Fusion layer are connected to the linear layer after the residual connection.

[0015] Among them, the first CINO model and the second CINO model are used for the input of Uyghur text and Chinese text respectively.

[0016] Furthermore, the first CINO model and the second CINO model convert the input Uyghur text and Chinese text into multi-dimensional vectors, including the Uyghur coding vector ug_vec and the Chinese coding vector zh_vec, respectively. The concat layer concatenates the Uyghur coding vector ug_vec and the Chinese coding vector zh_vec and inputs them into the Fusion layer. After the Fusion layer aligns the Uyghur and Chinese vectors in the high-dimensional space, the output of the Fusion layer is added to the input and then input into the linear layer for consistency judgment, and the judgment result is output.

[0017] Furthermore, the first CINO model and the second CINO model consist of 12 layers of transformers.

[0018] Furthermore, the Fusion layer consists of a multi-layer feed-forward neural network FFN.

[0019] Furthermore, the linear layer performs binary classification through the sigmoid function to obtain consistent or inconsistent judgment results.

[0020] Furthermore, the judgment model is trained through samples, and the sample construction process is as follows:

[0021] Obtain the conversation text between the question-answering robot and the Uyghur credit review user during the credit review process, as well as the application information submitted by the credit review user;

[0022] The Uyghur text of the answers given by the credit review user to the questions is extracted from the conversation text, and the Chinese text of the answers to the same questions corresponding to the application information is extracted. The Uyghur text and Chinese text of the answers to the same questions are used as features. If the semantics of the Uyghur text and the Chinese text are consistent, the semantic consistency is used as the label to complete the construction of a positive sample. If the semantics of the Uyghur text and the Chinese text are inconsistent, the semantic inconsistency is used as the label to complete the construction of a negative sample.

[0023] Furthermore, the question-answering robot will output a pre-set question list to the credit review user in sequence. The questions in the question list are set according to the application information submitted when applying for a loan.

[0024] The present invention is implemented as follows: a Uyghur language consistency determination system in letter review, the system comprising:

[0025] The acquisition unit, preprocessing unit, trained judgment model and output unit are connected in sequence;

[0026] The collection unit is used to collect the Uyghur voice answers given by the credit review user to the current question of the question-answering robot.

[0027] The preprocessing unit converts the collected Uyghur voice answers into Uyghur text, and reads the Chinese text corresponding to the answer to the current question of the question-answering robot from the application information. The Uyghur text and Chinese text corresponding to the answer to the current question of the question-answering robot are input into the trained judgment model. The trained judgment model determines whether the semantics of the Uyghur text and the Chinese text are the same, and outputs the judgment result through the output unit, and the judgment result is consistent or inconsistent.

[0028] The present invention uses a judgment model formed based on deep learning technology to directly perform semantic consistency recognition on input Uyghur text and Chinese text, without the need to translate Uyghur into Chinese first and then perform consistency judgment. This greatly improves the consistency recognition speed and accuracy, and can perform semantic consistency recognition of Uyghur and Chinese online. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A flow chart of a method for determining consistency of Uyghur language in a letter review provided by an embodiment of the present invention;

[0030] Figure 2 A schematic diagram of the structure of a determination model provided by an embodiment of the present invention;

[0031] Figure 3 A schematic diagram of the structure of a Uyghur language consistency determination system in a letter review provided by an embodiment of the present invention. Implementation

[0032] The specific implementation modes of the present invention are further explained in detail below by describing the embodiments with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.

[0033] Figure 1 This is a flow chart of a method for determining consistency of Uyghur language in a letter review provided by an embodiment of the present invention. The method is specifically as follows:

[0034] (1) The question-answering robot outputs the current question to the credit review user, collects the Uyghur voice answer from the credit review user, and converts the Uyghur voice answer content into Uyghur text;

[0035] In an embodiment of the present invention, the question-answering robot outputs a pre-set question list to the credit review user in sequence. The questions in the question list are set according to the application information submitted when applying for a loan. The question-answering robot extracts the question text from the question list in sequence, converts the question text into voice data, and outputs it to the credit review user. After receiving the answer from the credit review user, the question robot extracts the next question text from the question list in sequence.

[0036] (2) Obtain the Chinese text of the answer to the current question in the application information submitted by the credit review user;

[0037] In an embodiment of the present invention, the information type of the application information submitted by the user is basically fixed. Therefore, questions are set according to the application information, and the positions of the corresponding answers to the relevant questions in the application form are determined. After obtaining the Uyghur answer of the credit review user, the Chinese text of the relevant answer is directly read from the corresponding position in the application form. It should be noted that when the credit review user makes a loan request, he submits the application information based on the application form given by the lender.

[0038] (3) The Uyghur text and the Chinese text corresponding to the current question are input into the trained judgment model. The judgment model outputs whether the current Uyghur text is semantically consistent with the Chinese text, and outputs a judgment result, which is consistent or inconsistent.

[0039] The present invention uses a dual-tower structure, one side is the encoding layer for Uyghur text, and the other side is the encoding layer for Chinese text. The vectors output by the two encoding layers are then simultaneously input into the Fusion layer for feature fusion. Finally, the result of the feature fusion layer is input into the consistency judgment layer, and finally the final consistent or inconsistent result is obtained.

[0040] Figure 2 The structural diagram of the determination model provided in the embodiment of the present invention only shows the part related to the embodiment of the present invention for the convenience of explanation. The system includes:

[0041] The first CINO model and the second CINO model;

[0042] A Concat layer connected to the first CINO model and the second CINO model;

[0043] Fusion layer connected to Concat layer;

[0044] The input and output of the Fusion layer are connected to the linear layer after the residual connection.

[0045] Among them, the first CINO model and the second CINO model are used for the input of Uyghur text and Chinese text respectively. The first CINO model and the second CINO model convert the input Uyghur text and Chinese text into multi-dimensional vectors respectively, including the Uyghur coding vector ug_vec and the Chinese coding vector zh_vec. The concat layer concatenates the Uyghur coding vector ug_vec and the Chinese coding vector zh_vec and then inputs them into the Fusion layer; after the Fusion layer aligns the Uyghur and Chinese vectors in the high-dimensional space, the output of the Fusion layer is added to the input, and then input into the linear layer for consistency judgment, and the judgment result is output.

[0046] The first CINO model and the second CINO model are composed of 12 layers of transformers. The first CINO model encodes the input Uyghur text and converts the input Uyghur text into a Uyghur encoding vector ug_vec. The second CINO model encodes the input Chinese text and converts the input Chinese text into a Chinese encoding vector zh_vec. The Uyghur encoding vector ug_vec and the Chinese encoding vector zh_vec are 768-dimensional vectors.

[0047] In an embodiment of the present invention, the fusion layer is composed of a multi-layer feedforward neural network (FFN), and the input of the fusion layer is the concatenation of the Uyghur coding vector ug_vec and the Chinese coding vector zh_vec, fusion_input = concat (ug_vec, zh_vec). The fusion layer aligns the Uyghur and Chinese vectors in high-dimensional space through two layers of FFN networks.

[0048] The main function of the linear layer is to classify the fused high-dimensional vectors output by the Fusion layer, and perform binary classification through the sigmoid function to obtain consistent or inconsistent judgment results. The input end of the linear layer uses a residual link, and the input includes the sum of the input and output of the Fusion layer.

[0049] In the embodiment of the present invention, the judgment model needs to be trained with samples before use. Only after the judgment model training is completed, the trained judgment model can be used for consistency judgment between the Uyghur answer and the Chinese answer in the letter review process. The sample construction process in the training process is specifically as follows:

[0050] Obtain the conversation text between the question-answering robot and the Uyghur credit review user during the credit review process, as well as the application information submitted by the credit review user;

[0051] The Uyghur text of the answers given by the credit review user to the questions is extracted from the conversation text, and the Chinese text of the answers to the same questions corresponding to the application information is extracted. The Uyghur text and Chinese text of the answers to the same questions are used as features. If the semantics of the Uyghur text and the Chinese text are consistent, the semantic consistency is used as the label to complete the construction of a positive sample. If the semantics of the Uyghur text and the Chinese text are inconsistent, the semantic inconsistency is used as the label to complete the construction of a negative sample.

[0052] Figure 3 The structural diagram of the Uyghur language consistency determination system in the letter review provided by the embodiment of the present invention, for the convenience of explanation, only the part related to the embodiment of the present invention is shown, and the system includes:

[0053] The acquisition unit, preprocessing unit, trained judgment model and output unit are connected in sequence;

[0054] The collection unit is used to collect the Uyghur voice answers given by the credit review user to the current question of the question-answering robot.

[0055] The preprocessing unit converts the collected Uyghur voice answers into Uyghur text, and reads the Chinese text corresponding to the answer to the current question of the question-answering robot from the application information. The Uyghur text and Chinese text corresponding to the answer to the current question of the question-answering robot are input into the trained judgment model. The trained judgment model determines whether the semantics of the Uyghur text and the Chinese text are the same, and outputs the judgment result through the output unit, and the judgment result is consistent or inconsistent.

[0056] The present invention uses a judgment model formed based on deep learning technology to directly identify the semantic consistency of input Uyghur text and Chinese text, without the need to translate Uyghur into Chinese first and then make a consistency judgment. This greatly improves the consistency recognition speed and accuracy, and can perform online semantic consistency recognition of Uyghur and Chinese.

[0057] The present invention has been described exemplarily. Obviously, the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the method concept and technical solution of the present invention, or the concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A method for determining consistency of Uyghur language in letter review, characterized in that: The method comprises the following steps: (1) The question-answering robot outputs the current question to the credit review user, collects the Uyghur voice answer from the credit review user, and converts the Uyghur voice answer content into Uyghur text; (2) Obtain the Chinese text of the answer to the current question in the application information submitted by the credit review user; (3) The Uyghur text and the Chinese text corresponding to the current question are input into the trained judgment model. The judgment model outputs whether the current Uyghur text is semantically consistent with the Chinese text and outputs a judgment result.

2. The method for determining consistency of Uyghur language in letter review as claimed in claim 1, characterized in that: The determination model comprises: The first CINO model and the second CINO model; A Concat layer connected to the first CINO model and the second CINO model; Fusion layer connected to Concat layer; The input and output of the Fusion layer are connected to the linear layer after the residual connection. Among them, the first CINO model and the second CINO model are used for the input of Uyghur text and Chinese text respectively.

3. The method for determining consistency of Uyghur language in letter review as claimed in claim 2, characterized in that: The first CINO model and the second CINO model convert the input Uyghur text and Chinese text into multi-dimensional vectors, including the Uyghur coding vector ug_vec and the Chinese coding vector zh_vec, respectively. The concat layer concatenates the Uyghur coding vector ug_vec and the Chinese coding vector zh_vec and inputs them into the Fusion layer. After the Fusion layer aligns the Uyghur and Chinese vectors in high-dimensional space, the output of the Fusion layer is added to the input and then input into the linear layer for consistency judgment, and the judgment result is output.

4. The method for determining consistency of Uyghur language in letter review as claimed in claim 2, characterized in that: The first CINO model and the second CINO model consist of 12 layers of transformers.

5. The method for determining consistency of Uyghur language in letter review as claimed in claim 2, characterized in that: The fusion layer consists of a multi-layer feed-forward neural network FFN.

6. The method for determining consistency of Uyghur language in letter review as claimed in claim 2, characterized in that: The linear layer performs binary classification through the sigmoid function to obtain consistent or inconsistent judgment results.

7. The method for determining consistency of Uyghur language in letter review as claimed in claim 1, characterized in that: The judgment model is trained through samples, and the sample construction process is as follows: Obtain the conversation text between the question-answering robot and the Uyghur credit review user during the credit review process, as well as the application information submitted by the credit review user; The Uyghur text of the answers given by the credit review user to the questions is extracted from the conversation text, and the Chinese text of the answers to the same questions corresponding to the application information is extracted. The Uyghur text and Chinese text of the answers to the same questions are used as features. If the semantics of the Uyghur text and the Chinese text are consistent, the semantic consistency is used as the label to complete the construction of a positive sample. If the semantics of the Uyghur text and the Chinese text are inconsistent, the semantic inconsistency is used as the label to complete the construction of a negative sample.

8. The method for determining consistency of Uyghur language in letter review as claimed in claim 1, characterized in that: The question-answering robot will output the pre-set question list to the credit review user in sequence. The questions in the question list are set based on the application information submitted when applying for a loan.

9. A Uyghur language consistency determination system in letter review, characterized in that: The system comprises: The acquisition unit, preprocessing unit, trained judgment model and output unit are connected in sequence; The collection unit is used to collect the Uyghur voice answers given by the credit review user to the current question of the question-answering robot. The preprocessing unit converts the collected Uyghur voice answers into Uyghur text, and reads the Chinese text corresponding to the answer to the current question of the question-answering robot from the application information. The Uyghur text and Chinese text corresponding to the answer to the current question of the question-answering robot are input into the trained judgment model. The trained judgment model determines whether the semantics of the Uyghur text and the Chinese text are the same, and outputs the judgment result through the output unit, and the judgment result is consistent or inconsistent.

Images