An NLP-based intelligent unification method for asset management product valuation table

Abstract

The present application relates to the field of financial technology artificial intelligence, and in particular to a method for intelligently unifying asset management product valuation tables based on NLP, which pre-processes original valuation table data, extracts full-text information of underlying accounting subject codes as input for an NLP model, trains a bert classification model, and thereby establishes a corresponding relationship between the original valuation table accounting subject system and the standard accounting subject system, for completing the cleaning of the valuation table. The method can be embedded in asset management product analysis software, thereby greatly reducing the workload of valuation table data analysis and improving data accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of financial technology artificial intelligence, in particular to a method for intelligent unification of asset management product valuation table based on NLP. BACKGROUND

[0002] In recent years, bank wealth management and asset management industry have developed very rapidly. At present, the scale of bank wealth management has exceeded 30 trillion. With the requirements of new asset management rules and bank wealth management net worth, bank wealth management and other asset management institutions need to manage hundreds of asset management products of different channels uniformly, form various regulatory reports, business reports and interface risk control systems. All business scenarios of investment management platform depend on unified and accurate valuation table data. The unification of data is the core of all business analysis. However, the valuation methods and valuation tables of different managers are different, and the accounting subjects are quite different. At present, there is a lack of a unified system to manage the valuation tables of different channels. Bank wealth management subsidiaries and other asset management institutions currently mainly adopt manual processing and artificial checking to integrate valuation table data, which consumes a lot of manpower and time, and the results often have errors.

[0003] Due to the frequent changes of valuation methods and subject systems of different managers, it is difficult to develop software through traversing business logic to process valuation table data. If the valuation system of a certain channel is slightly changed, the previously developed system may cause errors. For the above reasons, it is difficult for each bank wealth management subsidiary and financial technology company to solve the problem of inconsistent product valuation table data and non-unified accounting subjects through system development based on the existing software development method. SUMMARY

[0004] In view of the defects in the prior art, the purpose of the present application is to provide a method for intelligent unification of asset management product valuation table based on NLP, which automatically unifies different types of valuation tables adopted by different managers into valuation tables under a standard subject system. NLP mentioned in this application is a general term in the field of artificial intelligence, which translates to natural language processing.

[0005] To achieve the above purpose, the technical solution adopted by the present application is:

[0006] A method for intelligent unification of asset management product valuation table based on NLP, comprising the following steps:

[0007] S1, data acquisition: sorting the historical valuation tables of different channels into the database, and acquiring or constructing a unified standard subject system, the standard subject system comprising a standard accounting subject system and a standard asset classification system;

[0008] S2, data preprocessing:

[0009] S21, tagging each asset management plan managed by each manager with the corresponding manager;

[0010] S22, using an algorithm to find the underlying accounting subject code of the original valuation table and cutting according to the subject code level, and splicing the underlying accounting subject name corresponding to the underlying accounting subject code according to the subject name corresponding to the code level;

[0011] S23, manually establishing the unique correspondence between the underlying accounting subject code of the original valuation table and the underlying accounting subject code of the standard subject system, taking the underlying accounting subject code of the standard subject system as the classification tag, and taking the underlying accounting subject name of the original valuation table after splicing in step S22 as the feature quantity, to construct a sample set for model training;

[0012] S3, model training: establishing an NLP model, and training the NLP model using the sample set generated in step S2;

[0013] S4, model optimization: using the NLP model trained in step S3 to convert the original valuation table into a valuation table under the standard subject system, manually checking the asset items and liability items of the valuation table before and after conversion, and marking the items with correct classification tags for optimization training of the NLP model;

[0014] S5, using the model optimized in step S4 to convert the original valuation table into a valuation table under the standard subject system.

[0015] Further, the algorithm in step S2 is: sorting the accounting subject code and the accounting subject name of each manager according to the length of the accounting subject code from small to large; traversing each accounting subject code A, using the accounting subject code A to cut all accounting subject codes C longer than him from left to right, if it can be cut, the cut accounting subject code C is expressed as A+ separator + B, where B is the remaining part code after cutting by A, if it cannot be cut, the accounting subject code that cannot be cut remains unchanged, if all accounting subjects longer than A code cannot be cut, A is marked as the underlying accounting subject code; extracting the underlying accounting subject code and the corresponding underlying accounting subject name, and splicing the underlying accounting subject name according to the corresponding underlying accounting subject code, the splicing method is: manager name + asset unit corresponding subject name + underlying asset corresponding subject name;

[0016] Further, the structure of the NLP model in step S3 includes one pre-training bert layer and one fully connected layer.

[0017] Further, the classification tag in step S23 is set as the standard asset classification category, and the trained model is used for asset classification of the valuation table items.

[0018] The NLP-based asset management product valuation table intelligent unification method has the following beneficial effects:

[0019] The method establishes an NLP model from historical valuation table data, and uses the learned model to unify different types of valuation tables into valuation tables under a standard subject system, thereby completing data cleaning of the valuation table. The method can be embedded in asset management product analysis software, thereby greatly reducing the workload of valuation table data analysis and improving data accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0020] The present application has the following drawings:

[0021] Figure 1 NLP model diagram of the present application;

[0022] Figure 2 Model training loss curve diagram of the present application;

[0023] Figure 3 Model training accuracy curve diagram of the present application. DETAILED DESCRIPTION

[0024] The present application will be further described in detail below in combination with the drawings.

[0025] The NLP-based asset management product valuation table intelligent unification method includes the following steps:

[0026] S1, data acquisition: organize historical valuation tables from different channels into a database, and obtain or construct a unified standard subject system, which includes a standard accounting subject system and a standard asset classification system.

[0027] To facilitate understanding of the present application, a sample table is provided, and the original valuation table sample table is shown in Table 1, and the valuation table under the standard accounting subject system is shown in Table 2.

[0028] Table 1

[0029] Manager name Account code Account name XX Fund 1002 Bank deposit XX Fund 100201 Bank deposit_current XX Fund 10020101 Bank deposit_current XX Fund 1002010101 Bank deposit_current_bank deposit XX Fund 10020102 Accrued interest on bank deposit XX Fund 1002010201 Interest receivable_bank deposit XX Fund 1102 Trading stock investment XX Fund 110201 XX Market_Ordinary shares already listed XX Fund 11020101 XX Market_Ordinary shares already listed_cost XX Fund 110201016AAAAA AA shares XX Fund 110201016BBBBB BB shares XX Fund 110233 XX Market_Ordinary shares already listed XX Fund 11023301 XX Market_Ordinary shares already listed_cost XX Fund 110233010CCCCC CC shares XX Fund 110233010DDDDD DD shares XX Fund 110234 XX Market_Ordinary shares already listed_Gem board XX Fund 11023401 XX Market_Ordinary shares already listed_Gem board_cost XX Fund 110234013EEEEE EE shares XX Fund 110234013FFFFF FF shares XX Fund 110283 XX Market_Ordinary shares already listed

[0030] Table 2

[0031]

[0032]

[0033] S2, data preprocessing:

[0034] S21, label each asset management plan managed by each manager with the corresponding manager;

[0035] S22, find the underlying accounting subject code of the original valuation table by using the algorithm, and cut according to the subject code level, and splice the underlying accounting subject name corresponding to the underlying subject code according to the subject name corresponding to the code level;

[0036] S23, manually establish the unique correspondence between the underlying accounting subject code of the original valuation table and the underlying accounting subject code of the standard subject system, take the underlying accounting subject code of the standard subject system as the classification label, and take the underlying accounting subject name of the original valuation table after splicing in step S22 as the feature quantity, and construct a sample set for model training.

[0037] S3, model training: establish an NLP model, and train the NLP model by using the sample set generated in step S2.

[0038] S4, model optimization: using the NLP model trained in step S3, convert the original valuation table into the valuation table under the standard subject system, manually check the asset items and liability items of the valuation table before and after the conversion, mark the items with correct classification labels for optimization training of the NLP model. It should be noted that the asset and liability items in the valuation table are the most core part, and ensuring the accuracy of the two basically ensures the accuracy of the model.

[0039] S5, using the model optimized in step S4 to convert the original valuation table into the valuation table under the standard subject system.

[0040] The algorithm described in step S2 is as follows: Sort the two fields of accounting subject code and accounting subject name for each manager in ascending order of accounting subject code length; traverse each accounting subject code A, and use accounting subject code A to cut all accounting subject codes C that are longer than it from left to right. If it can be cut, the cut accounting subject code C is represented as A + separator + B, where B is the remaining code after being cut by A. If it cannot be cut, the accounting subject codes that cannot be cut remain unchanged. If all accounting subjects with a code length longer than A cannot be cut, then A is marked as the underlying accounting subject code; extract the underlying accounting subject code and the corresponding underlying accounting subject name, and concatenate the underlying accounting subject name according to the corresponding underlying accounting subject code. The concatenation method is: manager name + asset unit one corresponding subject name + asset unit two corresponding subject name + ... + underlying asset corresponding subject name. The delimiter can be a period, a space, or other characters that serve a separating function. The delimited accounting code ABCD...X represents the combination of asset unit 1 (A), asset unit 2 (AB), asset unit 3 (ABC), asset unit 4 (ABCD), and X, and so on. Each asset unit can correspond to a unique account name. For example, the account code "1101" corresponds to the account name "Trading Stock Investment"; the account code "1101.01" corresponds to the account name "XX Exchange_Listed_Stock"; the original account "110101CCCC XX Shares" is delimited into "1101 Trading Stock Investment", "1101.01XX Exchange_Listed_Stock", and "1101.01.CCCC XX Shares", where "1101.01.CCCC" is the underlying asset code. The concatenated accounting account corresponding to the underlying asset code 1101.01.CCCC is: "Manager Name" + "Trading Stock Investment" + "XX Exchange_Listed_Stock" + "XX Shares".

[0041] The NLP model described in step S3 consists of a pre-trained BERT layer and a fully connected layer. Specifically, after concatenating the original accounting subject names, some accounting subjects are labeled based on a new classification standard (such as the standard accounting subject system). The labeled text set is then divided into a training set, a validation set, and a test set according to a 60% to 20% ratio. Figure 1 As shown, a classification BERT model is constructed. The part that needs to be trained is to match the original phrases with standard accounting subject labels. The labels are digitally encoded to form a multi-classification task. The model structure consists of a pre-trained BERT layer and a fully connected layer. The Harbin Institute of Technology pre-trained model "chinese-bert-wwm-ext" is used. The maximum string length is set to 128, the batch size is 4, and the learning rate is set to 2e-5.Figure 1 AA, BB, CC, DD represent the input original accounting subject splicing string, such as "manager name" + "trading stock investment" + "XX listed stock" + "XX stock"; the accounting subject classification can be the standard accounting subject bottom classification of the target manager or the asset classification, and only the classification needs to be encoded with numbers, and the loss function curve in training is shown in Figure 2 When the model is applied, the corresponding valuation table accounting subject code and accounting subject name of each manager are input, and the corresponding model is used to convert the accounting subject code and name under the new standard to complete the reclassification of the assets. The reading ability part of the bert model is pre-trained and very accurate, so that the accuracy can be close to 100%, and due to the existence of the verification mechanism, the accuracy can be 100%.

[0042] Figure 2 The NLP model training loss curve of the application is shown in Figure Figure 3 The NLP model training precision curve of the application is shown in Figure. It can be seen that with the increase of the training round, the loss decreases and the precision improves, and the model converges.

[0043] The method of the application can be calculated by historical valuation table modeling to quickly and accurately unify the valuation tables of different sources into the accounting subject system and asset classification system stipulated by the bank wealth management subsidiary and other asset management institutions, so as to complete the data cleaning of the valuation table. The cleaned valuation table sample table is shown in Table 3.

[0044] Table 3

[0045]

[0046]

[0047] The contents not described in detail in the specification belong to the prior art known to those skilled in the art.

Claims

1. A method for intelligent unification of asset management product valuation tables based on NLP, characterized in that, Includes the following steps: S1. Data Acquisition: Organize and store historical valuation tables from different channels into a database, and acquire or construct a unified standard account system, which includes a standard accounting account system and a standard asset classification system; S2, Data Preprocessing: S21. Label each asset management plan managed by a manager with the corresponding manager tag; S22. Use an algorithm to find the underlying accounting subject codes of the original valuation table and cut them according to the subject code hierarchy. Then, concatenate the underlying accounting subject names corresponding to the underlying accounting subject codes according to the subject names corresponding to the code hierarchy. S23. Manually establish a unique correspondence between the underlying accounting subject codes of the original valuation table and the underlying accounting subject codes of the standard subject system. Use the underlying accounting subject codes of the standard subject system as classification labels and the underlying accounting subject names of the original valuation table spliced ​​in step S22 as feature quantities to construct a sample set for model training. S3. Model Training: Build an NLP model and train the NLP model using the sample set generated in step S2; S4. Model Optimization: Using the NLP model trained in step S3, the original valuation table is transformed into a valuation table under the standard subject system. The asset and liability items of the valuation table before and after the transformation are manually checked. Items with incorrect transformation are given the correct classification labels and used for the optimization training of the NLP model. S5. Use the optimized model from step S4 to transform the original valuation table into a valuation table under the standard subject system. The algorithm described in step S2 is as follows: sort the two fields of accounting subject code and accounting subject name for each manager in ascending order of accounting subject code length; Iterate through each accounting subject code A. Using accounting subject code A, cut all accounting subject codes C longer than it from left to right. If it can be cut, the cut accounting subject code C is represented as A + separator + B, where B is the remaining code after being cut by A. If it cannot be cut, the accounting subject codes that cannot be cut remain unchanged. If all accounting subjects longer than A cannot be cut, then A is marked as the underlying accounting subject code. Extract the underlying accounting subject code and the corresponding underlying accounting subject name. Concatenate the underlying accounting subject name according to the corresponding underlying accounting subject code. The concatenation method is: manager name + asset unit corresponding subject name + underlying asset corresponding subject name.

2. The intelligent unification method for asset management product valuation tables based on NLP as described in claim 1, characterized in that, The structure of the NLP model described in step S3 includes a pre-trained BERT layer and a fully connected layer.

3. The intelligent unification method for asset management product valuation tables based on NLP as described in claim 1, characterized in that, The classification labels described in step S23 are set as standard asset classification categories, and the trained model is used to classify assets in the valuation table.

Images