A loan purpose text classification method and system based on a convolutional neural network

Abstract

The application discloses a loan purpose text classification method based on a convolutional neural network, which comprises the following steps: S1, establishing a green loan identification text classification model based on a text convolutional neural network, wherein the green loan identification text classification model based on the text convolutional neural network is an end-to-end text classification model, which directly learns relevant features of a green loan from original text and identifies green loan purpose text based on the relevant features; and S2, identifying whether a loan purpose text is a green loan according to the green loan identification text classification model based on the text convolutional neural network. Corresponding systems, electronic devices and computer readable storage media are also disclosed. By constructing an end-to-end green loan identification text classification model based on TextCNN, the relevant features of a green loan are directly learned from original text, and efficient and accurate identification of green loan purpose text is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer science and financial technology, and particularly relates to a loan purpose text classification method and system based on a convolutional neural network (CNN), and particularly relates to a text classification model for green loan identification established by using a convolutional neural network technology, so as to perform loan purpose text classification based on the model. BACKGROUND

[0002] With the improvement of environmental protection consciousness and the development of green economy, green loans, as an important financial tool to support environmental protection projects and sustainable development projects, have important significance for risk control and policy making of financial institutions. However, the traditional green loan identification method mainly relies on an expert system or a rule-based text analysis method, and the traditional green loan identification method is artificial identification, that is, artificial labeling classification (a total of 212 classifications including 211 classifications and non-green loans) is performed on each loan purpose in combination with the Green Industry Guidance Catalogue 2019 issued by the state. These methods usually require a large amount of manpower for feature extraction and rule making, which not only consumes time and effort, but also is difficult to adapt to the diversity and complexity of text data.

[0003] In actual application, the description text of the loan purpose often contains rich semantic information, and the traditional text classification method is difficult to effectively capture and process these information, resulting in low recognition accuracy. In addition, with the increase in the number of green loan projects, how to quickly and accurately process a large amount of text data has become a challenge for financial institutions. Although the existing machine learning method improves the automation degree of classification to a certain extent, it is still limited by the quality of feature engineering and the generalization ability of the model, so that it becomes a problem to be solved to automatically, efficiently and accurately complete this process. SUMMARY

[0004] In order to solve the problems in the prior art, the application provides a loan purpose text classification method and system based on a convolutional neural network (CNN), which directly learns relevant features of green loans from original text by constructing an end-to-end green loan identification text classification model based on TextCNN, and realizes efficient and accurate identification of green loan purpose text, wherein the green loan identification text classification model based on TextCNN is a deep learning model, which automatically extracts key features in the text through a convolutional layer, does not require complex feature engineering, and can effectively handle the high dimensionality and complexity of text data. Thus, the deficiencies of the prior art in processing green loan identification tasks are overcome, and the limitations of the prior art in processing green loan identification tasks are overcome, especially the deficiencies in the classification accuracy of traditional models, the generalization ability of models, and the efficiency of processing large-scale text data are made up. The powerful capabilities of deep learning technology are used to improve the accuracy and efficiency of green loan identification, provide a new technical support for financial institutions, and promote the development of green finance.

[0005] The application provides a loan purpose text classification method based on a convolutional neural network, which comprises the following steps:

[0006] S1, a green loan identification text classification model based on a text convolutional neural network is established, wherein the green loan identification text classification model based on the text convolutional neural network is an end-to-end text classification model, which directly learns relevant features of green loans from original text, and identifies green loan purpose text based on the relevant features;

[0007] S2, whether the loan purpose text is a green loan is identified based on the green loan identification text classification model based on the text convolutional neural network.

[0008] Preferably, the S1 comprises the following steps:

[0009] S11, loan data is collected, and a data set required for establishing the green loan identification text classification model based on the text convolutional neural network is generated based on the loan data;

[0010] S12, a text convolutional neural network model is trained based on the training data set to automatically extract features capable of identifying green loans in the simulated loan purpose text; wherein the text convolutional neural network model is an NLP model;

[0011] S13, fusing the NLP model and the strategy model to form a final green loan identification text classification model based on a text convolutional neural network, and formulating specific green recognition rules according to an output result of the final green loan identification text classification model based on the text convolutional neural network.

[0012] Preferably, the S11 comprises:

[0013] (1) collecting actual loan data from banks and regulatory agencies;

[0014] (2) generating relevant loan data based on a generative artificial intelligence large model relying on its own learning and reasoning ability and referring to the Green Industry Guidance Catalogue 2019;

[0015] (3) constructing a simulated loan purpose text data set based on the actual loan data and the relevant loan data;

[0016] (4) manually labeling the simulated loan purpose text to establish a labeled data set.

[0017] Preferably, the S12 comprises:

[0018] (1) data cleaning: cleaning all data in the labeled data set, including removing duplicate records, correcting incorrect information, and handling missing values, to ensure data quality; then, performing word segmentation processing on the simulated loan purpose text, and mapping the text into a numerical vector using Sogou and a pre-trained word vector, i.e., converting it into a numerical feature, so that the model can learn;

[0019] (2) performing data mining based on all data in the cleaned labeled data set and data fed back by regulatory agencies and banks to generate a green recognition keyword and non-green keyword list;

[0020] (3) setting two basic layer architectures of the text convolutional neural network model to realize the entire green recognition model, the two basic layer architectures comprising a text classification layer and a strategy layer, wherein the text classification layer is composed of an NLP text classification model to perform preliminary classification of the simulated loan purpose text for labeling and green recognition; the strategy layer is composed of a strategy model based on a set access rule, including a positive strategy layer and a negative strategy layer, for accurate labeling and green recognition of the simulated loan purpose text, and the link of the entire green recognition model is to first perform text classification by the NLP model, and then apply a strategy to accurately fine-tune the output result of the NLP model through the strategy layer;

[0021] (4) training the green loan identification text classification model based on the text convolutional neural network based on the cleaned labeled data set to generate the NLP model; and generating the strategy model based on the green recognition keyword and non-green keyword list.

[0022] Preferably, the training of the green loan identification text classification model based on the text convolutional neural network based on the cleaned labeled data set and the text classification layer in the two basic layer architectures generates the NLP model, which comprises:

[0023] Dataset splitting: the labeled data set is divided into a training set and a test set, wherein the training set is used for model learning, and the test set is used for evaluating the generalization ability and performance of the model;

[0024] Model training: training the green loan identification text classification model based on the text convolutional neural network based on the data in the training set, wherein the green loan identification text classification model based on the text convolutional neural network automatically extracts text features through the convolutional layer, reduces the dimension through the pooling layer, and finally classifies through the fully connected layer and the output layer; In the training process, the weights of the model are continuously adjusted through the back propagation algorithm to minimize the loss function;

[0025] Model evaluation: based on the data in the test set, the performance of the green loan identification text classification model based on the text convolutional neural network is evaluated, and the evaluation is based on the accuracy, recall rate, F1 score, and topn classification standard accuracy and green identification accuracy. The indicators measure the classification effect of the green loan identification text classification model based on the text convolutional neural network;

[0026] The strategy model is generated based on the green recognition keyword and non-green keyword list and the strategy layer in the two basic layer architectures, which comprises:

[0027] The two levels required for building the strategy model are the green recognition strategy layer and the non-green strategy layer;

[0028] The two levels required for building the strategy model based on the keyword strategy method and the access strategy method;

[0029] The green recognition strategy layer corresponds to the green recognition strategy, which includes two kinds of green recognition keywords for green identification; wherein the first kind of green recognition keyword is derived from real bank and regulatory data; the second kind of green recognition keyword is derived from the documents in the national standard;

[0030] The non-green strategy layer corresponds to the non-green strategy, which includes two kinds; wherein the first non-green strategy is: if the length of the text is less than or equal to four after removing the punctuation marks, it is directly determined as non-green; the second non-green strategy is the non-green keyword strategy, wherein the non-green keywords are mainly derived from the loan data provided by the bank and the regulatory agency. The mode that frequently appears in multiple loan records is retained as a non-green strategy, and the loan purpose that meets the mode will be identified as non-green.

[0031] Preferably, the S13 fuses the NLP model and the strategy model to form a final green loan identification text classification model based on a text convolutional neural network, and the S13 comprises the following steps of:

[0032] A model fusion rule is set: the NLP model and the strategy model output two groups of classification results respectively, a possibility classification column of covering the NLP model result by the strategy model result is used, if the strategy model classification result is null, the classification result of the NLP model is reserved; if the strategy model classification result is not null, the classification result of the strategy model is reserved.

[0033] Preferably, the S2 comprises the following steps of:

[0034] S21, application development is performed, an interface for calling by a backend system is created, a trained model is packaged as a green identification model service, and seamless integration with an existing loan processing system is realized;

[0035] S22, real-time loan data of the loan processing system is acquired and converted into loan purpose text;

[0036] S23, a green loan identification text classification model based on a text convolutional neural network is formed to form an identification link, and whether the loan purpose text is a green loan is identified, and the S23 comprises the following steps of:

[0037] (1) a loan purpose is reported by a front end and stored in a database, and an id corresponding to the loan purpose is transmitted to a model interface by a back end;

[0038] (2) data corresponding to the loan purpose in the database is acquired as model input, a classification result corresponding to each loan purpose is output by the model, and then the classification result is written into the database;

[0039] (3) the classification result in the database is read by the back end, the classification result is returned to the front end, and the whole identification link is completed

[0040] A second aspect of the application provides a loan purpose text classification system based on a convolutional neural network, and the loan purpose text classification system comprises:

[0041] A model establishment module is configured to establish a green loan identification text classification model based on a text convolutional neural network, wherein the green loan identification text classification model based on the text convolutional neural network is an end-to-end text classification model, learns relevant features of a green loan directly from original text, and identifies green loan purpose text based on the relevant features;

[0042] A loan purpose text classification module is configured to identify whether loan purpose text is a green loan based on the green loan identification text classification model based on a text convolutional neural network.

[0043] The third aspect of the present application provides an electronic device comprising a processor and a memory, the memory storing a plurality of instructions, and the processor being configured to read the instructions and perform the method according to the first aspect.

[0044] The fourth aspect of the present application provides a computer-readable storage medium storing a plurality of instructions, the plurality of instructions being readable by a processor and executable to perform the method according to the first aspect.

[0045] The method, system and electronic device provided by the present application have the following beneficial effects:

[0046] By constructing an end-to-end green loan identification text classification model based on TextCNN, the relevant features of green loans are learned directly from the original text, and efficient and accurate identification of green loan purpose text is realized. The green loan identification text classification model based on TextCNN is a deep learning model that automatically extracts key features in the text through convolutional layers, without complex feature engineering, and can effectively handle the high dimensionality and complexity of text data. Thus, the deficiencies of the prior art in processing green loan identification tasks are overcome, and the limitations of the prior art in processing green loan identification tasks are overcome, especially the deficiencies in the classification accuracy and model generalization ability of traditional models and the efficiency of processing large-scale text data are overcome. The powerful capabilities of deep learning technology are utilized to improve the accuracy and efficiency of green loan identification, providing a new technical support for financial institutions and promoting the development of green finance. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 A flowchart of the process of establishing a green loan identification text classification model based on a text convolutional neural network in the loan purpose text classification method based on a convolutional neural network according to the present application.

[0048] Figure 2 A principle flowchart of the process of establishing a green loan identification text classification model based on a text convolutional neural network in the loan purpose text classification method based on a convolutional neural network according to the present application.

[0049] Figure 3 A flowchart of the process of identifying whether a loan purpose text is a green loan according to the green loan identification text classification model based on a text convolutional neural network according to the present application.

[0050] Figure 4 A structural diagram of the electronic device according to the present application. DETAILED DESCRIPTION

[0051] For better understanding of the above technical solutions, the above technical solutions will be described in detail below in combination with the drawings of the specification and specific embodiments.

[0052] The embodiment provides a loan purpose text classification method based on a convolutional neural network, which comprises the following steps:

[0053] S1, a green loan identification text classification model based on a text convolutional neural network (TextCNN) is established, wherein the green loan identification text classification model based on the text convolutional neural network (TextCNN) is an end-to-end text classification model, which directly learns relevant features of a green loan from original text and identifies green loan purpose text based on the relevant features;

[0054] S2, whether the loan purpose text is a green loan is identified based on the green loan identification text classification model based on the text convolutional neural network.

[0055] As shown in Figures 1-2 , as a preferred embodiment, the S1 comprises:

[0056] S11, loan data is collected, and a data set required for establishing the green loan identification text classification model based on the text convolutional neural network is generated based on the loan data;

[0057] In the embodiment, the S11 comprises:

[0058] (1) actual loan data from banks and regulatory agencies is collected;

[0059] As a data source 1 of the embodiment: bank data and regulatory data

[0060] The first data source obtains relevant data from banks and regulatory agencies. The data is divided into two types: green loan data and non-green loan data. The green loan data is mainly obtained from regulatory agencies, and the non-green data is mainly obtained from banks supervised by regulatory agencies. After the data obtained in this channel is cleaned and de-duplicated, it is given to business personnel for manual labeling. According to the national standard "Green Industry Guidance Catalogue 2019", the above loan purposes will be classified into 212 categories, of which the non-green data is a separate category, and the remaining green data is subdivided into 211 categories.

[0061] (2) relevant loan data generated by a generative artificial intelligence large model based on its learning and reasoning ability and referring to "Green Industry Guidance Catalogue 2019";

[0062] As data source 2 of the embodiment: AI generated data

[0063] The second data source is related loan data generated by the generative artificial intelligence large model relying on its learning and reasoning ability and referring to the Green Industry Guidance Catalogue 2019

[0064] The way the large model generates data is as follows:

[0065] (A) Create a large model instance on the Ali Cloud Refining Large Model platform and upload the Green Industry Guidance Catalogue 2019 to the enterprise knowledge base, so that the subsequent large model can read the content therein;

[0066] (B) Obtain the api interface of the large model instance and write the corresponding promot prompt to call the large model to generate green loan data. The specific promot is as follows:

[0067] ”'

[0068] Promote: You are a professional green loan identification personnel who is very familiar with the content of the Green Industry Guidance Catalogue (2019 Edition) document. You are very familiar with green loan uses and their corresponding classifications

[0069] Now you need to do the green loan data generation work and generate 15 loan purpose data. Remember: Now these purpose data belong to the {dic} directory. The following is the loan purpose of the industry and meets the content of the Green Industry Guidance Catalogue (2019 Edition) document. The text data generated in the text data contains national standards and is summarized into one

[0070] The result is returned in json format, the format is {{"loan purpose", [all generated purposes are placed in this list]}}. That is, a key and a list. The loan purpose you generated is text data and does not need braces

[0071] ”';

[0072] (C) Write a program to call the large model interface to generate data. The dic above is a variableized directory name. The program is filled in a loop. Each time a certain number of data under a certain classification directory is obtained, it is saved to the local;

[0073] (D) Finally, the generated data is subjected to data quality exploration and cleaning, such as removing null values, removing duplicate values, removing redundant index values, and manually verifying whether the generated data content meets the provisions in the Green Industry Guidance Catalogue (2019 Edition).

[0074] (3) Based on the actual loan data and the related loan data, construct a simulated loan purpose text data set;

[0075] In this embodiment, the data obtained in this step is as follows:

[0076]

[0077] (4) Artificially annotating the simulated loan purpose text to establish an annotated data set.

[0078] S12, training a text convolutional neural network model based on the training data set to automatically extract features in the simulated loan purpose text that can be used for green loan identification; wherein the text convolutional neural network model is an NLP model.

[0079] In this embodiment, the NLP model is trained using the annotated data set obtained in S11. The NLP model has strong generalization ability and can recognize synonyms. For example, if the text learned by the NLP model is "buy steel for green building" and the corresponding classification is A, then if the subsequent text encountered is "buy metal for green construction", the NLP model can still classify it into the A category. The NLP model classifies in the form of probability, and it outputs the probability of each classification after recognizing each text. By selecting the top n probabilities, the corresponding directory standard of the text can be captured, and the flexibility is high. Furthermore, this embodiment can make a keyword strategy by mining effective keywords in the data, and then adjust the results of the NLP model.

[0080] As a preferred embodiment, S12 comprises:

[0081] (1) Data cleaning: all data in the annotated data set is cleaned to ensure data quality and provide accurate input for model training;

[0082] In this embodiment, the collected loan data is cleaned, including removing duplicate records, correcting incorrect information, and handling missing values, to ensure data quality. Subsequently, the simulated loan purpose text is segmented, and the text is mapped into a numerical vector using Sogou and a pre-trained word vector, i.e., converted into a numerical feature, so that the model can learn

[0083] (2) Based on all data in the cleaned annotated data set and data feedback from regulatory agencies and banks, artificial data mining is performed to generate green and non-green keyword lists;

[0084] (3) Set two basic layer architectures of the text convolutional neural network model to realize the entire green model, the two basic layer architectures include: a text classification layer and a strategy layer, wherein the text classification layer is composed of an NLP text classification model to perform preliminary classification of the simulated loan purpose text for benchmarking and green identification; the strategy layer is composed of a strategy model, which is constructed based on a set of access rules, including a positive strategy layer and a negative strategy layer, for performing accurate benchmarking and green identification of the simulated loan purpose text, wherein the text classification layer and the strategy layer complement each other, the strategy layer serves as a powerful supplement to the natural language processing (NLP) model, and together with the text classification layer, the strategy layer constitutes a complete green identification model, the link of the entire green model is that the NLP model performs text classification first, and then the strategy layer applies a strategy to accurately fine-tune the output result of the NLP model to improve the accuracy of the final green loan identification;

[0085] In this embodiment, the network structure of TextCNN mainly includes the following levels:

[0086] Input layer: receives the word vector representation of the input text, which is usually converted into a fixed-dimensional vector by a pre-trained word embedding model.

[0087] Convolutional layer: uses multiple convolution kernels (filters) of different sizes to extract local features in the text

[0088] Activation layer: usually connected after the convolutional layer, uses a nonlinear activation function (such as ReLU) to increase the nonlinear representation ability of the model.

[0089] Pooling layer: used to reduce the feature dimension, reduce the amount of calculation, and extract the main features. Common pooling operations include max pooling and average pooling.

[0090] Fully connected layer: maps the pooled features to one or more fully connected layers for nonlinear transformation, preparing for the classification task.

[0091] Output layer: this selection selects the class with the largest selected value as the first prediction result

[0092] The specific model construction structure is:

[0093] (A) 3 convolutional layers corresponding to (2, 3, 4) three different convolution kernel sizes.

[0094] (B) Each convolutional layer uses nn.Conv2d, the input channel number is 1 (because the text data is one-dimensional, so the channel number is set to 1), the output channel number is 256 here, the height of the convolution kernel is the word embedding dimension, and the width is the convolution kernel size (2, 3, 4).

[0095] (C) dropout layer to prevent overfitting parameter settings are 0.5;

[0096] (D) the output of the fully connected layer is 212 corresponding to the 212 classifications of green loans;

[0097] The remaining parameters are:

[0098] The number of training rounds (11 rounds);

[0099] The number of samples per batch (128);

[0100] The length of text padding (210);

[0101] The learning rate (1e-3);

[0102] The dimension of the embedding layer (300).

[0103] (4) Based on the cleaned labeled data set, the green loan recognition text classification model based on the text convolutional neural network is trained to generate the NLP model; and a strategy model is generated based on the green key word and non-green key word word table.

[0104] In this embodiment, the green loan recognition text classification model based on the text convolutional neural network is trained based on the cleaned labeled data set and the text classification layer in the two basic layer architectures in step (4) to generate the NLP model, which includes:

[0105] Dataset splitting: the labeled data set is divided into a training set and a test set (in this embodiment, the training set and the test set are divided according to the proportions of 80% and 20%), wherein the training set is used for model learning, and the test set is used for evaluating the generalization ability and performance of the model;

[0106] Model training: based on the data in the training set, the green loan recognition text classification model based on the text convolutional neural network is trained, wherein the green loan recognition text classification model based on the text convolutional neural network automatically extracts text features through a convolutional layer, reduces the dimension through a pooling layer, and finally classifies through a fully connected layer and an output layer; during the training process, the weights of the model are constantly adjusted through a backpropagation algorithm to minimize the loss function.

[0107] Model evaluation: based on the data in the test set, the performance of the green loan recognition text classification model based on the text convolutional neural network is evaluated, and the evaluation is based on the accuracy, recall rate, F1 score, and topn classification standard accuracy and green identification accuracy to measure the classification effect of the green loan recognition text classification model based on the text convolutional neural network.

[0108] In this embodiment, the step (4) of generating a strategy model based on the green-recognized keywords and non-green-recognized keywords table and the strategy layer in the two basic layers includes:

[0109] The two layers required for constructing the strategy model are the green-recognized strategy layer and the non-green-recognized strategy layer.

[0110] The two layers required for constructing the strategy model based on the keyword strategy method and the access strategy method; wherein the keyword strategy method refers to a logical judgment procedure of "if the loan purpose contains the word a, then it is classified into category b", the keyword is derived from the data provided by the bank and the regulatory authority, valuable keywords are mined by business personnel, and technical personnel make each module in the strategy model in the form of regular expressions; the access strategy method is a one-vote-no-data strategy, for example, if the text length of the loan purpose is less than four characters, it is directly recognized as non-green.

[0111] As a preferred embodiment, the green-recognized strategy layer corresponds to the green-recognized strategy, which includes being recognized as green based on two green-recognized keywords; wherein the first green-recognized keyword is derived from real bank and regulatory data, for example, the two texts "waste alkali transportation" and "waste acid transportation" have the same pattern, then the language pattern 'waste? transportation' is extracted as a strategy, the loan purpose that meets the pattern will be labeled as 2.2.3 hazardous waste transportation and recognized as green; the second green-recognized keyword is derived from the documents in the national standard, such as the standard name in the "Green Industry Guidance Catalogue (2019 Edition)", for example, "comply with Boiler Air Pollutant Emission Standard (GB 13271)", considering that part of the green investment products and links should meet relevant standards, therefore the strategy item will search whether the loan purpose contains this standard name, if it contains, it will be directly recognized as green and accurately labeled.

[0112] Examples of green keywords:

[0113]

[0114]

[0115] As a preferred embodiment, the non-green strategy layer corresponds to a non-green strategy, and the non-green strategy includes two; wherein the first non-green strategy is: if the length of the text is less than or equal to four after removing punctuation, it is directly determined as non-green; theoretically, a text of less than four characters may be neutral, but it is more likely that it cannot prove itself to be a green loan, and the actual data analysis results are also the same. Therefore, this rule strategy is of great value, with a small sacrifice to obtain a large green identification ability, for example, 998 of the 1000 texts are non-green, and only two are green, so that the risk of washing green is reduced at a small cost, and at the same time, the two texts that are mistakenly killed have a high probability of being on the correct track by the green recognition strategy layer. The second non-green strategy is the non-green keyword strategy, wherein the non-green keywords are mainly derived from the loan data provided by the bank and the regulatory authority, such as the loan data "The 100,000 yuan loan is used to repay someone's debt", and the text pattern "repay.*?Debt" hidden in this text frequently appears in multiple loan records, so the pattern is retained as a non-green strategy. The loan purpose that meets the pattern is determined as non-green as follows: part of the non-green keywords.

[0116] Non-green keyword examples:

[0117]

[0118] S13, the NLP model and the strategy model are fused to form a final green loan recognition text classification model based on a text convolutional neural network, and specific green recognition rules are formulated according to the output results of the final green loan recognition text classification model based on the text convolutional neural network.

[0119] As a preferred embodiment, the S13 of the NLP model and the strategy model are fused to form a final green loan recognition text classification model based on a text convolutional neural network, and specific green recognition rules are formulated according to the output results of the final green loan recognition text classification model based on the text convolutional neural network.

[0120] Set the model fusion rule: the NLP model and the strategy model will output two sets of classification results respectively, so two model fusion methods need to be defined.

[0121] NLP model output example

[0122] In the classification of loan purposes in the Green Industry Guidance Catalogue 2019, there are 211 categories plus the non-green category, and the final classification result has a total of 212 categories. After any loan purpose is input into the NLP model, the model will output all the classification orders, that is, the 212 categories are sorted by possibility. In this invention, only the results of the first two classifications are considered to construct the model, because the first two classifications have the highest possibility. Here, the first three classification examples are shown.

[0123] For example, the NLP model output result example:

[0124]

[0125] Strategy model result example

[0126] The strategy model only outputs one classification result due to the uniqueness of the mapping between keywords and classification results.

[0127]

[0128] Model fusion rule

[0129] Because the strategy model result is highly interpretable and is formulated by human intervention, the strategy model result is used as the criterion for model construction, which is greater than the NLP model result.

[0130] Fusion method: Finally, use the strategy model result to cover the possibility classification 1 column of the NLP model result; the rule is to keep the NLP classification result if the strategy model classification result is empty; if the strategy model classification result is not empty, keep the strategy model classification result.

[0131] Final NLP model and strategy model integrated fusion output structure example:

[0132]

[0133] As a preferred embodiment, the data input into the green loan identification text classification model based on the text convolutional neural network and the final classified category directory example (the model outputs 212 classification results, arranged in order of probability, and only the classification directory with the highest probability is displayed here):

[0134]

[0135] For example, if the loan purpose is "purchase of environmental monitoring and control system, etc.", the model will return the classification directory of this purpose, i.e. "6.4.5 Enterprise Environmental Monitoring", and it will be identified as a green loan. For the purpose of "purchase of injection molding machine", the model identifies it as the classification directory corresponding to non-green loan, i.e. "non-green".

[0136] As a preferred embodiment, the S13 formulates specific green recognition rules according to the output results of the final green loan identification text classification model based on the text convolutional neural network, which includes:

[0137] If possible classification 1 is not green, the loan is identified as a non-green client and only the "non-green" is displayed. Otherwise, the possible classification 1 is output. At this time, if possible classification 2 is not non-green, possible classification 2 is also output. If possible classification 2 is non-green, only possible classification 1 is output.

[0138] Display example of the present embodiment

[0139] For example, for "solar energy utilization facility construction and operation" in the above table, the client output display is

[0140] ["3.2.2 Solar energy utilization facility construction and operation", "5.4.3 Urban integrated integrated energy supply facility construction and operation"] for "purchase of photovoltaic power generation equipment" only displays ["3.2.2 Solar energy utilization facility construction and operation"] for "the loan of 1.5 million yuan is used to purchase cement" displays ["non-green"].

[0141] Referring to Figure 3 , as a preferred embodiment, the S2 comprises:

[0142] S21, application development is carried out, an interface is created for the call of the back-end system, the trained model is packaged as a green identification model service, and seamless integration with the existing loan processing system is realized;

[0143] S22, real-time loan data of the loan processing system is obtained and converted into loan purpose text;

[0144] S23, a recognition link is formed based on the green loan identification text classification model based on the text convolutional neural network, and identification of whether the loan purpose text is a green loan is carried out, comprising:

[0145] (1) the front-end fills in the loan purpose and stores it in the database, and the back-end transmits the id corresponding to the loan purpose to the model interface;

[0146] (2) the model queries the data corresponding to the loan purpose in the database as the model input, the model outputs the classification result corresponding to each loan purpose, and then writes the classification result into the database;

[0147] (3) the back-end reads the classification result in the database, returns the classification result to the front-end, and completes the entire recognition link

[0148] The present embodiment also provides a loan purpose text classification system based on a convolutional neural network, comprising:

[0149] The model establishing module is configured to establish a green loan identification text classification model based on a text convolutional neural network (TextCNN).

[0150] The loan purpose text classification module is configured to identify whether the loan purpose text is a green loan based on the green loan identification text classification model based on the text convolutional neural network.

[0151] The application further provides a memory storing a plurality of instructions for implementing the method of the first embodiment.

[0152] As shown in Figure 4 The application further provides an electronic device including a processor 301 and a memory 302 connected to the processor 301, wherein the memory 302 stores a plurality of instructions, and the instructions can be loaded and executed by the processor to enable the processor to perform the method of the first embodiment.

[0153] Although the preferred embodiments of the application have been described, those skilled in the art who understand the basic inventive concept can make additional changes and modifications to the embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the application. Obviously, those skilled in the art can make various modifications and variations to the application without departing from the spirit and scope of the application. Thus, if these modifications and variations of the application fall within the scope of the claims of the application and their equivalents, the application also intends to include these modifications and variations.

Claims

1. A text classification method for loan purpose based on convolutional neural networks, characterized in that, include: S1. Establish a text classification model for green loan recognition based on text convolutional neural networks. The text classification model for green loan recognition based on text convolutional neural networks is an end-to-end text classification model that directly learns the relevant features of green loans from the original text and identifies the text on the purpose of green loans based on the relevant features. S2, Based on the text classification model for green loan identification based on text convolutional neural network, identify whether the loan purpose text is a green loan; S1 includes: S11, Collect loan data and generate a dataset based on the loan data to build the text classification model for green loan recognition based on text convolutional neural network; S12, a text convolutional neural network model is trained based on the training dataset to automatically extract features from simulated loan purpose text that can be used to identify green loans; wherein, the text convolutional neural network model is an NLP model; S13, the NLP model and the policy model are fused to form the final text classification model for green loan recognition based on text convolutional neural network, and specific green loan recognition rules are formulated based on the output of the final text classification model for green loan recognition based on text convolutional neural network. S11 includes: (1) Collect actual loan data from banks and regulatory agencies; the actual loan data includes bank data and regulatory data. The data is divided into two types: green loan data and non-green loan data. Green loan data comes from regulatory agencies, and non-green data comes from banks regulated by regulatory agencies. The data obtained through this channel is cleaned and deduplicated before being handed over to business personnel for manual labeling. According to the national standard "Green Industry Guidance Catalogue (2019 Edition)", the above loan purposes will be divided into 212 categories, of which non-green data is counted as a separate category, and the remaining green data is subdivided into 211 categories. (2) Relevant loan data generated by a generative artificial intelligence large model based on its own learning and reasoning capabilities, referencing the "Green Industry Guidance Catalogue (2019 Edition)"; the methods by which the large model generates data include: A. Create a large model instance on the Alibaba Cloud Large Model Platform and upload the "Green Industry Guidance Catalogue (2019 Edition)" to the enterprise knowledge base so that the large model can read its contents in the future; B. Obtain the API interface of the large model instance and write the corresponding promotion prompt to call the large model to generate green loan data; C. Write a program to call the large model interface to generate data. The dic above is a variable directory name, which is filled in cyclically by the program. Each time a certain number of data under a certain category directory are obtained, they are saved locally. D. Finally, the generated data is cleaned and tested for quality, including removing null values, duplicate values, and redundant index values. The generated data content is manually checked to see if it complies with the provisions of the "Green Industry Guidance Catalogue (2019 Edition)". (3) Construct a text dataset simulating loan usage based on the actual loan data and the relevant loan data; (4) Manually annotate the simulated loan purpose text to establish an annotated dataset; S12 includes: (1) Data cleaning: Clean all data in the labeled dataset, including removing duplicate records, correcting error information, and handling missing values ​​to ensure data quality; then, perform word segmentation on the simulated loan purpose text, and use Sogou and pre-trained word vectors to map the text into numerical vectors, that is, transform it into numerical features so that the model can learn. (2) Based on all the data in the cleaned labeled dataset and the data fed back by regulatory agencies and banks, manual data mining was performed to generate a list of green keywords and non-green keywords; (3) The text convolutional neural network model is set up with two basic layer architectures to realize the entire green identification model. The two basic layer architectures include: a text classification layer and a strategy layer. The text classification layer is composed of an NLP text classification model to perform preliminary classification and identification of simulated loan purpose texts. The strategy layer is composed of the strategy model, which is built based on the set admission rules and includes a positive strategy layer and a negative strategy layer for precise identification of simulated loan purpose texts. The text classification layer and the strategy layer complement each other. The strategy layer is a powerful supplement to the natural language processing NLP model and together with the text classification layer, it constitutes a complete green identification model. The entire green identification model chain is that the NLP model first performs text classification, and then the strategy layer applies a strategy to the output of the NLP model for precise fine-tuning. (4) The NLP model is generated by training the text classification model for green loan recognition based on the cleaned labeled dataset and the text classification layer in the two basic layer architectures; a strategy model is generated based on the vocabulary of green loan recognition keywords and non-green loan keywords and the strategy layer in the two basic layer architectures. The NLP model is generated by training the text classification model for green loan recognition based on the cleaned labeled dataset and the text classification layer in the two basic layer architectures. Dataset splitting: The labeled dataset is divided into a training set and a test set, wherein the training set is used for model learning and the test set is used to evaluate the model's generalization ability and performance; Model Training: The text classification model for green loan recognition based on text convolutional neural networks is trained based on the data in the training set. The text classification model for green loan recognition based on text convolutional neural networks automatically extracts text features through convolutional layers, reduces dimensionality through pooling layers, and finally performs classification through fully connected layers and output layers. During the training process, the model weights are continuously adjusted through the backpropagation algorithm to minimize the loss function. Model Evaluation: The performance of the text classification model for green loan identification based on the text convolutional neural network is evaluated based on the data in the test set. The evaluation measures the classification effect of the text classification model for green loan identification based on the text convolutional neural network by using the metrics of accuracy, recall, F1 score, and the calculation of top n classification accuracy and green identification accuracy. The strategy model generated based on the vocabulary of green-recognition keywords and non-green keywords, and the strategy layer in the two basic layer architectures, includes: The strategy model is constructed using two layers: a green-recognition strategy layer and a non-green strategy layer. The strategy model is constructed based on two layers: a keyword strategy method and an access strategy method. The keyword strategy method refers to a logical judgment procedure that states "if the loan purpose contains the word 'a', then it is classified under category 'b'". The keywords are derived from data provided by banks and regulatory agencies. Business personnel discover valuable keywords from the data, and technical personnel use regular expressions to create various modules in the strategy model. The access strategy method is a veto strategy. The green recognition strategy layer corresponds to the green recognition strategy, which includes identifying green based on two types of green recognition keywords; the first type of green recognition keyword comes from real bank and regulatory data; the second type of green recognition keyword comes from documents in national standards. The non-green strategy layer corresponds to the non-green strategy, which includes two types. The first non-green strategy is: if the length of the text after removing punctuation is less than or equal to four, it is directly defined as non-green. The second non-green strategy is the non-green keyword strategy, where non-green keywords come from loan data provided by banks and regulatory agencies. Patterns that frequently appear in multiple loan records are retained as non-green strategies, and loan purposes that conform to this pattern will be identified as non-green. The step S13, which involves fusing the NLP model with the policy model to form the final text classification model for green loan recognition based on a text convolutional neural network, includes: Define the model fusion rules: The NLP model and the strategy model will each output two sets of classification results, thereby defining the way the two models are fused; the strategy model results are used to cover the probability classification columns of the NLP model results. If the strategy model classification result is null, the NLP model classification result is retained; if the strategy model classification result is not null, the strategy model classification result is retained. S2 includes: S21 involves application-oriented development, creating interfaces that can be called by backend systems, encapsulating the trained model into a green recognition model service, and achieving seamless integration with the existing loan processing system. S22, Obtain real-time loan data from the loan processing system and convert it into loan purpose text; S23, based on the text classification model for green loan identification using the text convolutional neural network, a recognition link is formed to identify whether the loan purpose text is a green loan, including: (1) The front end fills in the loan purpose and enters it into the database, and the back end sends the ID corresponding to the loan purpose to the model interface; (2) The model queries the data corresponding to the loan purpose in the database as the model input, the model outputs the classification result corresponding to each loan purpose, and then writes the classification result into the database; (3) The backend reads the classification results from the database and returns the classification results to the frontend to complete the entire recognition chain.

2. A loan purpose text classification system based on convolutional neural networks, used to implement the method of claim 1, characterized in that, include: The model building module is used to build a text classification model for green loan recognition based on a text convolutional neural network. The text classification model for green loan recognition based on a text convolutional neural network is an end-to-end text classification model that directly learns the relevant features of green loans from the original text and identifies the text on the purpose of green loans based on the relevant features. The loan purpose text classification module is used to identify whether a loan purpose text is a green loan based on the text convolutional neural network-based green loan identification text classification model.

3. An electronic device, characterized in that, It includes a processor and a memory, the memory storing multiple instructions, and the processor for reading the instructions and executing the method as described in claim 1.

4. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions, which can be read by a processor and executed as described in claim 1.

Images