Intelligent Recognition Method and System for Green Finance Attributes Based on OCR and NLP Technologies

Abstract

The present invention provides an intelligent determination method and system for green financial attributes based on OCR and NLP technologies, including: Step S1: Extract and recognize image information to obtain unstructured character data information; Step S2: Perform data structuring construction on the unstructured character data information to obtain standardized structured character data; Step S3: Preprocess the standardized structured character data to obtain data in the form of word vectors or matrices; Step S4: Input the text word vectors into a fully connected network for analysis to obtain specific attribute classification results. The present invention can efficiently analyze the document information of relevant materials and give valuable reference indicators; the present invention provides a fast and efficient financial attribute determination service for bank staff; the present invention uses OCR and NLP technologies to improve the review efficiency of green financial loans for bank customers.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and system for intelligently identifying green financial attributes based on OCR and NLP technologies. Background Art

[0002] In 2020, senior central bank officials emphasized the need to develop green financial technology at a relevant conference, noting that developing financial technology "through the application of digital technologies such as big data, artificial intelligence, the Internet of Things, cloud computing, and blockchain" can effectively alleviate the pain points and difficulties of related businesses. Green finance, as a bridge, can more effectively guide more social capital into green industries. Therefore, enhancing the ability to identify and categorize the attributes of green finance is particularly important.

[0003] At present, most green finance attribute identification systems use the form of keyword dictionaries. This identification method often results in inaccurate segmentation of special words in the financial professional field. To solve this problem, a large amount of manual work is required to set up domain-specific dictionaries and word priorities. Manual intervention will inevitably cause subjective differences in the attribute identification system.

[0004] Patent document CN115048940A (application number: CN202210724689.7) discloses a Chinese financial text data enhancement method based on entity word attribute features and back translation. The method includes: first, performing text preprocessing on the input financial text; then, using named entity recognition and vocabulary matching methods, replacing entity words in the financial text that are more difficult to translate with referents to generate the intermediate text of the financial text and a referent-entity mapping relationship dictionary; then, using the back translation method to enhance the intermediate text of the financial text, and using the referent-entity mapping relationship dictionary to restore the referents in the text, thereby generating a new text with similar semantics to the original text. However, this invention does not utilize OCR and NLP technologies. Summary of the invention

[0005] In view of the defects in the existing technology, the purpose of the present invention is to provide a method and system for intelligent identification of green financial attributes based on OCR and NLP technology.

[0006] According to the present invention, a method for intelligently identifying green financial attributes based on OCR and NLP technologies includes:

[0007] Step S1: extract and recognize image information to obtain unstructured character data information;

[0008] Step S2: structuring the unstructured character data to obtain standardized structured character data;

[0009] Step S3: Preprocess the standardized structured character data to obtain data in the form of word vectors or matrices;

[0010] Step S4: Input the text word vectors into a fully connected network for analysis to obtain specific attribute classification results.

[0011] Preferably, in the said step S1:

[0012] The data provided by the user includes structured data and unstructured data; among which, for the processing of unstructured data, an auxiliary detection model needs to be added to pre-detect the unstructured image data, structurally restore each detected box discriminated by the model, and convert it into template data with a structural relationship and input it into the extraction and recognition model at the backend together with the structured data;

[0013] In the data extraction and recognition stage, perform character recognition on each keyword field, and recognize each image character as the corresponding computer character; apply a recognition neural network model in the character recognition to recognize the keyword field information in the case of being blocked by seals and watermarks and having low pixels, and the obtained unstructured character data information will be passed to the next module for data structuring.

[0014] Preferably, in the said step S2:

[0015] Temporarily store the output unstructured character data information in the server as the input-output stream of the data, waiting for the next data structuring process;

[0016] For the structured construction, use the method of registering the detection box and the character content in correspondence. The detection box output by the detection model outputs the relative coordinates of two diagonal points, and then input the content in the corresponding box into the corresponding coordinate position. Through an intelligent algorithm, align and adapt the detection box, and output the standardized structured character data of the original image, providing the standardized structured character data for the NLP model at the backend.

[0017] Preferably, in the said step S3:

[0018] Preprocessing of text data:

[0019] Use the NLP intelligent model to perform word analysis and named entity recognition on the standardized structured character data, clean the data to remove invalid characters, limit the maximum length of the input text, and if there is an exceeded part, use the slicing input method to obtain standard data with a consistent text length;

[0020] Word vector representation of the text:

[0021] Convert standard data with consistent text lengths to obtain data in the form of word vectors or matrices. Use a pre-trained word embedding model to tokenize the text into dense low-dimensional word vectors, and set the neuron inactivation coefficient to reduce overfitting. Design an end-to-end solution for information extraction of index items. Through data annotation of predefined index items, the text content with given index item labels is obtained by neural network training.

[0022] Preferably, in the step S4:

[0023] Input the text word vectors into a fully connected network for analysis to obtain specific attribute classification results. Select the Bi-LSTM model structure of Bert to process the text content with given index item labels, and then perform text classification through a specially designed fully connected layer and CRF model, and output the attribute classification results:

[0024] It includes the following two steps:

[0025] Step S4.1: Automatically extract features in the tokens from the text word vectors with given labels. Use the Word Embedding sequence of a complete text as the input for each time step of the bidirectional LSTM, and then concatenate the hidden state sequences output by the forward LSTM and the hidden states output by the backward LSTM at each position to obtain a complete hidden state sequence;

[0026] Step S4.2: Input the hidden state into the CRF layer for processing to achieve sequence prediction at the line text level. The CRF layer models the hidden state sequence in the text and calculates the transition probabilities of each state, and classifies the semantic information of the line text.

[0027] According to a green finance attribute intelligent recognition system based on OCR and NLP technologies provided by the present invention, it includes:

[0028] Module M1: Extract and recognize image information to obtain unstructured character data information;

[0029] Module M2: Perform data structuring construction on the unstructured character data information to obtain standardized structured character data;

[0030] Module M3: Preprocess the standardized structured character data into text data, and obtain data in the form of word vectors or matrices;

[0031] Module M4: Input the text word vectors into a fully connected network for analysis to obtain specific attribute classification results.

[0032] Preferably, in the module M1:

[0033] The data provided by the user includes structured data and unstructured data; among them, the processing of unstructured data requires adding an auxiliary detection model to pre-detect the unstructured image data, structurally restore each detection box discriminated by the model, and convert it into template data with a structural relationship and input it into the extraction and recognition model at the backend together with the structured data;

[0034] In the extraction and recognition stage of the data, text recognition is performed on each keyword field, and each image text is recognized as the corresponding computer character; the recognition neural network model is applied in the text recognition to recognize the keyword field information in the case of being blocked by seals and watermarks and having low pixels, and the obtained unstructured character data information will be passed to the next module for data structuring.

[0035] Preferably, in the module M2:

[0036] The output unstructured character data information is temporarily stored in the server as the input and output stream of the data, waiting for the next data structuring process;

[0037] The structured construction uses the method of registering the detection box and the character content in correspondence. The relative coordinates of two points at the diagonal of the detection box output by the detection model are output, and then the content in the corresponding box is input to the corresponding coordinate position. The detection box is aligned and adapted through an intelligent algorithm, and the standardized structured character data of the original image is output, providing the standardized structured character data for the NLP model at the backend.

[0038] Preferably, in the module M3:

[0039] Text data preprocessing:

[0040] Through the NLP intelligent model, word analysis and named entity recognition are performed on the standardized structured character data, invalid characters are removed from the character data by data cleaning, the maximum text length input is restricted, and if there is an exceeding part, it is input in a sliced manner to obtain standard data with a consistent text length;

[0041] Word vector representation of the text:

[0042] The standard data with a consistent text length is converted to obtain data in the form of word vectors or matrices. The pre-trained word embedding model is used to segment the text into dense low-dimensional word vectors, and the neuron inactivation coefficient is set to reduce the overfitting situation. An end-to-end solution for information extraction of the index item is designed. Through data annotation of the predefined index item, the text content with a given index item label is obtained by neural network training.

[0043] Preferably, in the module M4:

[0044] The text word vectors are input into a fully connected network for analysis to obtain specific attribute classification results. The Bi-LSTM model structure of Bert is selected to process the given index item label text content, and then through a specially designed fully connected layer and CRF model for text classification work, the attribute classification results are output:

[0045] It includes the following two steps:

[0046] Module M4.1: Automatically extract features in the word segmentation from the text word vectors of the given labels through the Bi-LSTM layer. Take the Word Embedding sequence of a complete text as the input of each time step of the bidirectional LSTM, and then splice the hidden state sequence output by the forward LSTM with the hidden states output by the reverse LSTM at each position to obtain a complete hidden state sequence;

[0047] Module M4.2: Input the hidden state into the CRF layer for processing to achieve sequence prediction at the line text level. The CRF layer models the hidden state sequence in the text and calculates the transition probabilities of each state to classify the semantic information of the line text.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. The present invention can efficiently analyze the document information of relevant materials and give valuable reference indicators;

[0050] 2. The present invention provides a fast and efficient financial attribute recognition service for bank staff;

[0051] 3. The present invention uses OCR and NLP technologies to improve the review efficiency of green financial loans for bank customers;

[0052] 4. The present invention provides a solution for dynamically adding new data and optimizing labels in the actual business scenario, which can realize the incremental training of the model and effectively reduce the iteration cycle after the model is launched;

[0053] 5. The present invention proposes a design idea for green financial attribute recognition and model iteration. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] By reading the following detailed description of the non-limiting embodiments with reference to the accompanying drawings, other features, objects and advantages of the present invention will become more apparent:

[0055] Figure 1 It is a business process flow chart of an intelligent green financial attribute recognition model. DETAILED DESCRIPTION OF THE INVENTION

[0056] The present invention will be described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that those of ordinary skill in the art can make several changes and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0057] Embodiment 1:

[0058] In the trend of big data, the present invention is based on new AI technologies such as optical character recognition and natural language processing, and significantly improves the discrimination ability of green financial attributes by developing and training relevant discrimination models.

[0059] Apply the optical character recognition model (Optical Character Recognition, OCR) and natural language processing model (Nature Language Processing, NLP) to conduct intelligent identification of the green financial attributes of the text materials of the loans granted. Through the intelligent algorithms of the optical character recognition model and natural language processing model, conduct green attribute identification on the financial materials to be reviewed submitted by users (including but not limited to project feasibility study reports, project feasibility study report approvals, project environmental impact assessment reports, project environmental assessment approvals, etc.). The present invention conducts intelligent recognition of the text information in the materials, extracts the theme keyword fields and keywords in the text, and then conducts natural language processing on the above keyword field information, and the neural network model determines whether the project has green financial related attributes.

[0060] According to a method for intelligent identification of green financial attributes based on OCR and NLP technologies provided by the present invention, as Figure 1 shown, it includes:

[0061] Step S1: Extract and recognize image information to obtain unstructured character data information;

[0062] Specifically, in the said step S1:

[0063] The data provided by the user includes structured data and unstructured data; among them, for the processing of unstructured data, an auxiliary detection model needs to be added to pre-detect the unstructured image data, structurally restore each detection box discriminated by the model, and convert it into template data with a structural relationship and input it into the extraction and recognition model at the backend together with the structured data;

[0064] In the data extraction and recognition stage, text recognition is performed on each keyword field, and each image text is recognized as a corresponding computer character. The recognition neural network model is applied in text recognition to recognize the keyword field information in the case of seal and watermark occlusion and low pixels. The obtained unstructured character data information will be passed to the next module for data structuring.

[0065] Step S2: Perform data structuring construction on the unstructured character data information to obtain standardized structured character data;

[0066] Specifically, in the step S2:

[0067] Temporarily store the output unstructured character data information in the server as the input and output stream of the data, waiting for the next data structuring process;

[0068] The structuring construction uses the method of registering the detection box and the character content in correspondence. The relative coordinates of two diagonal points of the detection box output by the detection model are output, and then the content in the corresponding box is input to the corresponding coordinate position. Through the intelligent algorithm, the detection box is aligned and adapted, and the standardized structured character data of the original image is output, providing the standardized structured character data for the NLP model at the back end.

[0069] Step S3: Preprocess the standardized structured character data to obtain data in the form of word vectors or matrices;

[0070] Specifically, in the step S3:

[0071] Text data preprocessing:

[0072] Through the NLP intelligent model, perform word and phrase analysis and named entity recognition on the standardized structured character data, clean the data to remove invalid characters, limit the maximum length of the input text, and use the slicing input method for the part exceeding the length to obtain standard data with consistent text lengths;

[0073] Word vector representation of the text:

[0074] Convert the standard data with consistent text lengths to obtain data in the form of word vectors or matrices. Use the pre-trained word embedding model to segment the text into dense low-dimensional word vectors, and set the neuron inactivation coefficient to reduce the overfitting situation. Design an end-to-end solution for information extraction of the index item. Through data annotation of the predefined index item, the text content with the given index item label is obtained by neural network training.

[0075] Step S4: Input the text word vectors into the fully connected network for analysis to obtain the specific attribute classification result.

[0076] Specifically, in the step S4:

[0077] The text word vectors are input into a fully connected network for analysis to obtain specific attribute classification results. The Bi-LSTM model structure of Bert is selected to process the given index item label text content, and then through a proprietary fully connected layer and a CRF model for text classification work, the attribute classification results are output:

[0078] It includes the following two steps:

[0079] Step S4.1: Automatically extract the features in the word segmentation from the text word vectors of the given labels. The Word Embedding sequence of a complete text is used as the input for each time step of the bidirectional LSTM, and then the hidden state sequence output by the forward LSTM is concatenated with the hidden states output by the backward LSTM at each position to obtain a complete hidden state sequence;

[0080] Step S4.2: The hidden state is passed into the CRF layer for processing to achieve sequence prediction at the line text level. The CRF layer models the hidden state sequence in the text and calculates the transition probabilities of each state to classify the semantic information of the line text.

[0081] Embodiment 2:

[0082] Embodiment 2 is a preferred example of Embodiment 1 to illustrate the present invention more specifically.

[0083] The present invention also provides a green finance attribute intelligent recognition system based on OCR and NLP technologies. The green finance attribute intelligent recognition system based on OCR and NLP technologies can be implemented by executing the process steps of the green finance attribute intelligent recognition method based on OCR and NLP technologies. That is, those skilled in the art can understand the green finance attribute intelligent recognition method based on OCR and NLP technologies as the preferred implementation manner of the green finance attribute intelligent recognition system based on OCR and NLP technologies.

[0084] According to a green finance attribute intelligent recognition system based on OCR and NLP technologies provided by the present invention, it includes:

[0085] Module M1: Extract and recognize image information to obtain unstructured character data information;

[0086] Specifically, in the module M1:

[0087] The data provided by the user includes structured data and unstructured data; among them, the processing of unstructured data requires adding an auxiliary detection model to pre-detect the unstructured image data, structurally restore each detected box discriminated by the model, and convert it into template data with a structural relationship, which is then input into the extraction and recognition model at the backend together with the structured data;

[0088] In the data extraction and recognition stage, text recognition is performed on each keyword field, and each image text is recognized as the corresponding computer character; a recognition neural network model is applied in the text recognition to recognize the keyword field information in the case of being blocked by seals and watermarks and having low pixels, and the obtained unstructured character data information will be passed to the next module for data structuring.

[0089] Module M2: Structurally construct the unstructured character data information to obtain standardized structured character data;

[0090] Specifically, in the module M2:

[0091] The output unstructured character data information is temporarily stored in the server as the input and output stream of the data, waiting for the next data structuring process;

[0092] The structural construction uses the method of corresponding registration of the detection box and the character content. The relative coordinates of the two diagonal points of the detection box output by the detection model are output, and then the content in the corresponding box is input to the corresponding coordinate position. The detection box is aligned and adapted through an intelligent algorithm, and the standardized structured character data of the original image is output, providing the standardized structured character data for the NLP model at the backend.

[0093] Module M3: Preprocess the standardized structured character data to obtain data in the form of word vectors or matrices;

[0094] Specifically, in the module M3:

[0095] Text data preprocessing:

[0096] Through the NLP intelligent model, word and phrase analysis and named entity recognition are performed on the standardized structured character data, invalid characters are removed from the character data through data cleaning, the maximum text length input is restricted, and if there is an excess part, it is input in a sliced manner to obtain standard data with a consistent text length;

[0097] Word vector representation of text:

[0098] Convert the standard data with consistent text lengths to data in the form of word vectors or matrices. Use a pre-trained word embedding model to tokenize the text into dense low-dimensional word vectors, and set the neuron inactivation coefficient to reduce overfitting. Design an end-to-end solution for information extraction of index items. Through data annotation of predefined index items, train the neural network to obtain the text content with given index item labels.

[0099] Module M4: Input the text word vectors into a fully connected network for analysis to obtain specific attribute classification results.

[0100] Specifically, in the module M4:

[0101] Input the text word vectors into a fully connected network for analysis to obtain specific attribute classification results. Select the Bi-LSTM model structure of Bert to process the text content with given index item labels, and then perform text classification through a proprietary fully connected layer and a CRF model, and output the attribute classification results:

[0102] It includes the following two steps:

[0103] Module M4.1: Automatically extract features in the tokens from the text word vectors with given labels. Use the Word Embedding sequence of a complete text as the input for each time step of the bidirectional LSTM, and then concatenate the hidden state sequences output by the forward LSTM with the hidden states output by the backward LSTM at each position to obtain a complete hidden state sequence;

[0104] Module M4.2: Input the hidden state into the CRF layer for processing to achieve sequence prediction at the line text level. The CRF layer models the hidden state sequence in the text and calculates the transition probabilities of each state to classify the semantic information of the line text.

[0105] Example 3:

[0106] Example 3 is a preferred example of Example 1 to more specifically illustrate the present invention.

[0107] A green finance attribute intelligent recognition method and system based on OCR and NLP technologies. This method is composed of a two-stage model of OCR and NLP. The main features of this model are:

[0108] Output of text character recognition and structured reconstruction of the first-stage OCR model;

[0109] Keyword extraction and attribute category intelligent recognition of the second-stage NLP model.

[0110] Specifically, it includes the following content:

[0111] (1) Image Information Extraction and Structured Data Construction

[0112] The present invention structurally restores the target detection frames obtained by the OCR intelligent algorithm, converts them into template data with structural relationships, and inputs them into the extraction and recognition model at the backend together with the structured data. In the character data recognition stage, it is necessary to perform text recognition on each keyword field and accurately recognize each image text as the corresponding character. The present invention applies the latest recognition neural network model in text recognition, which can accurately recognize keyword field information such as the title, table header, and body content of the text in the case of being blocked by seals and watermarks and having low pixels.

[0113] (2) Text Data Preprocessing and Word Vector Representation

[0114] The present invention performs word and phrase analysis and named entity recognition on the structured character data through the NLP intelligent model. And uses the pre-trained word embedding model to generate dense low-dimensional word vectors for each word and phrase in the text, and sets the neuron inactivation coefficient to reduce the overfitting of the model.

[0115] (3) Text Classification of the Fully Connected Layer

[0116] In this stage, the Bi-LSTM model structure of Bert is selected to process the to-be-processed word and phrase data in the word vector layer, and then the text classification work is carried out through a specially designed fully connected layer and CRF model. The Bi-LSTM layer can automatically extract the features in the word segmentation, and perform feature extraction on a complete text to obtain a complete hidden state sequence. The CRF layer performs text-level sequence prediction to judge whether the project belongs to the green finance project from various granularities.

[0117] (4) Establishment of the Green Finance Attribute Recognition System

[0118] The established AI intelligent recognition system is mainly divided into three parts: the client, the AI big data basic service platform, and the green finance system. The interaction is carried out through Web communication in the middle. Then the extracted sample data is input into the green finance system to provide the reference recognition result of the green finance attribute for the customer. Finally, the sample data is sorted out to provide data support for version iteration.

[0119] Example 4:

[0120] Example 4 is the preferred example of Example 1 to more specifically illustrate the present invention.

[0121] The present invention applies an OCR model and an NLP model to provide an intelligent identification system for the green finance attributes of the text materials of the loans granted, and the implementation methods mainly include information extraction and recognition of text images, construction of structured data, preprocessing of natural language text data, word vector representation of text, text classification of fully connected layers, establishment of a green finance attribute identification system, and version iteration of the new system.

[0122] By means of AI technologies such as OCR character recognition and NLP language processing, various types of original texts are processed to form structured data, and intelligent modules such as enterprise names, project names, loan flows, and project construction are refined, so as to provide users with convenient and efficient intelligent identification services for green finance attributes. The specific steps are as follows:

[0123] Step 1: Image information extraction and recognition.

[0124] First of all, the data provided by users mainly has two types: structured and unstructured (unstructured data: such as project environmental impact assessment approval documents in PDF or scanned pictures, etc.; structured data: such as enterprise loan application information registered in the credit system). Among them, the processing of unstructured data is relatively complex, and an auxiliary detection model needs to be added to pre-detect the unstructured image data. Then, the detection frames discriminated by the model are structurally restored, and after being converted into template data with a structural relationship, they are input into the extraction and recognition model at the backend together with the structured data. In the data extraction and recognition stage, it is necessary to perform character recognition on each key field, and accurately recognize each image character as the corresponding computer character. The present invention applies the latest recognition neural network model in character recognition, and can accurately recognize key field information such as titles, table headers, and body content in the case of being blocked by seals and watermarks and having low pixels. The obtained unstructured character data information will be passed to the next module for data structuring.

[0125] Step 2: Data structuring construction.

[0126] The unstructured character data information output by the OCR recognition model is temporarily stored in the server as the input and output stream of the data, waiting for the next data structuring process. If the unstructured character data is directly passed into the NLP model for lexical analysis and named entity recognition, there will be a situation of abnormal word segmentation. Therefore, it is necessary to perform structuring construction on the character data. The structuring construction of the present invention uses the method of corresponding registration of detection frames and character content. First, the relative coordinates of the two diagonal points of the detection frame output by the detection model are output, and then the content in the corresponding frame is input into the corresponding coordinate position. However, the sizes of different detection frames are different, and there will be a situation of misaligned text information. In the face of this problem, the present invention performs alignment and adaptation operations on the detection frames through intelligent algorithms. Finally, the standardized structured character data of the original image is output, providing the standardized structured character data for the NLP model at the backend;

[0127] Step 3: Preprocessing of text data.

[0128] The core work at this stage is to perform word analysis and named entity recognition on the standardized structured character data through the NLP intelligent model. First, there are a large number of invalid characters in the original character data transmitted by the superior. It is necessary to clean the character data, remove the invalid characters, limit the maximum input length of the text to 512, and use the slicing input method for the part exceeding the length to obtain standard data with consistent text lengths.

[0129] Step 4: Word vector representation of text.

[0130] Convert the standard data with consistent text lengths into data in a processable word vector or matrix form. The present invention uses a pre-trained word embedding model to convert text tokenization into dense low-dimensional word vectors (WordEmbedding), and sets the neuron inactivation coefficient to reduce the overfitting situation. Due to the existence of various semantic relationships (synonyms), context relationships between tokens in the corpus, and the complexity of word vector dimensions in text analysis. The present invention designs an end-to-end solution for information extraction of index items, and through data annotation of predefined index items, the text content with given index item labels is obtained by neural network training.

[0131] Step 5: Text classification by the fully connected layer.

[0132] Input the text word vectors of the previous stage into the fully connected network for analysis to obtain the specific attribute classification results. In this stage, the Bi-LSTM model structure of Bert is selected to process the text content with given index item labels, and then the text classification work is carried out through a specially designed fully connected layer and CRF model, and the final attribute classification results are output.

[0133] The said Step 5 includes the following two steps:

[0134] Step 5.1: Automatically extract the features in the tokens from the text word vectors with given labels. Take the Word Embedding sequence of a complete text as the input of each time step of the bidirectional LSTM, and then splice the hidden state sequence output by the forward LSTM with the hidden states output by the backward LSTM at each position to obtain a complete hidden state sequence;

[0135] Step 5.2: Input the hidden state into the CRF layer for processing to achieve sequence prediction at the line text level. The CRF layer models the hidden state sequence from t0 to t in the text and calculates the transition probabilities of each state, and finally gives the overall detailed semantic information classification of the line text. i

[0136] Those skilled in the art know that in addition to implementing the systems, devices, and their respective modules provided by the present invention in the form of pure computer-readable program code, it is entirely possible to logically program the method steps so that the systems, devices, and their respective modules provided by the present invention are implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, embedded microcontrollers, etc. to achieve the same program. Therefore, the systems, devices, and their respective modules provided by the present invention can be considered as a kind of hardware component, and the modules included therein for implementing various programs can also be regarded as the structures within the hardware component; the modules for implementing various functions can also be regarded as either software programs for implementing the method or the structures within the hardware component.

[0137] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. An intelligent identification method for green finance attributes based on OCR and NLP technologies, characterized in that, Including: Step S1: Extract and recognize image information to obtain unstructured character data information; Step S2: Perform data structuring construction on the unstructured character data information to obtain standardized structured character data; Step S3: Preprocess the standardized structured character data to obtain word vectors; Step S4: Input the text word vectors into a fully connected network for analysis to obtain specific attribute classification results; In the said Step S1: The data provided by the user includes structured data and unstructured data; among them, for the processing of unstructured data, an auxiliary detection model needs to be added to pre-detect the unstructured image data, and the detection boxes discriminated by the model are structurally restored, and after being converted into template data with a structural relationship, they are input into the extraction and recognition model at the backend together with the structured data; In the data extraction and recognition stage, text recognition is performed on each key field, and each image text is recognized as the corresponding computer character; in text recognition, a recognition neural network model is applied to recognize the key field information in the case of being blocked by seals and watermarks, and the obtained unstructured character data information will be passed to the next module for data structuring; In the said Step S3: Text data preprocessing: Use the NLP intelligent model to perform word analysis and named entity recognition on the standardized structured character data, clean the data to remove invalid characters, limit the input of the maximum text length, and if there is an exceeded part, use the slicing input method to obtain standard data with consistent text length; Word vector representation of text: Convert the standard data with consistent text length to obtain word vectors, use the pre-trained word embedding model to segment the text into dense low-dimensional word vectors, and set the neuron inactivation coefficient to reduce the overfitting situation, design an end-to-end solution for information extraction of index items, and through data annotation of the predefined index items, train the text content with the given index item labels by the neural network.

2. The intelligent identification method for green financial attributes based on OCR and NLP technologies according to claim 1, wherein In the said Step S2: Temporarily store the output unstructured character data information in the server as the input and output stream of the data, waiting for the next data structuring process; For structuring construction, use the method of corresponding registration of detection boxes and character content. The detection model outputs the relative coordinates of two diagonal points of the detection box, and then inputs the content in the corresponding box into the corresponding coordinate position. Through intelligent algorithms, the detection boxes are aligned and adapted, and the standardized structured character data of the original image is output, providing the standardized structured character data for the NLP model at the backend.

3. The intelligent identification method for green finance attributes based on OCR and NLP technologies according to claim 1, wherein, In the said Step S4: Input the text word vectors into a fully connected network for analysis to obtain specific attribute classification results. Select the Bi-LSTM model structure of Bert to process the given index item label text content, and then perform text classification work through a specially designed fully connected layer and CRF model, and output the attribute classification results: Including the following two steps: Step S4.1: Automatically extract the features in the word segmentation from the text word vectors of the given tags through the Bi-LSTM layer. Take the Word Embedding sequence of a complete text as the input of each time step of the bidirectional LSTM, and then concatenate the hidden state sequence output by the forward LSTM with the hidden states output by the backward LSTM at each position to obtain a complete hidden state sequence; Step S4.2: Pass the hidden state into the CRF layer for processing to achieve sequence prediction at the line text level. The CRF layer models the hidden state sequence in the text and calculates the transition probabilities of each state, and classifies the semantic information of the line text.

4. An intelligent identification system for green finance attributes based on OCR and NLP technologies, characterized in that, Including: Module M1: Extract and recognize image information to obtain unstructured character data information; Module M2: Construct data structuring for the unstructured character data information to obtain standardized structured character data; Module M3: Perform preprocessing on the standardized structured character data and obtain word vectors; Module M4: Input the text word vectors into a fully connected network for analysis to obtain specific attribute classification results; In the said Module M1: The data provided by the user includes structured data and unstructured data; among them, the processing of unstructured data requires adding an auxiliary detection model to pre-detect the unstructured image data, and restore the structured detection frames discriminated by the model, and then input them into the extraction and recognition model at the backend together with the structured data after being converted into template data with a structural relationship; In the extraction and recognition stage of the data, perform character recognition on each key field, and recognize each image character as the corresponding computer character; apply a recognition neural network model in the character recognition to recognize the key field information in the case of being blocked by seals and watermarks, and the obtained unstructured character data information will be passed to the next module for data structuring; In the said Module M3: Preprocessing of text data: Through the NLP intelligent model, perform word analysis and named entity recognition on the standardized structured character data, clean the character data to remove invalid characters, limit the input of the maximum text length, and if there is an excess part, use the slicing input method to obtain standard data with the same text length; Word vector representation of text: Convert the standard data with the same text length to obtain word vectors, use the pre-trained word embedding model to convert the text word segmentation into dense low-dimensional word vectors, and set the neuron inactivation coefficient to reduce the overfitting situation, design an end-to-end solution for information extraction of the index item, and through the data annotation of the predefined index item, obtain the text content with the given index item label by neural network training 5. The intelligent recognition system for green financial attributes based on OCR and NLP technologies according to claim 4, characterized in that, In the said Module M2: Temporarily store the output unstructured character data information in the server as the input and output stream of the data, waiting for the next data structuring process; The structured construction uses the method of corresponding registration of detection frames and character content. The relative coordinates of two diagonal points of the detection frame output by the detection model are output, and then the content in the corresponding frame is input to the corresponding coordinate position. Through an intelligent algorithm, the detection frame is aligned and adapted, and the standardized structured character data of the original image is output, providing the standardized structured character data for the NLP model at the back end.

6. The intelligent recognition system for green financial attributes based on OCR and NLP technologies according to claim 4, characterized in that, In the module M4: The text word vectors are input into a fully connected network for analysis to obtain specific attribute classification results. The Bi-LSTM model structure of Bert is selected to process the given metric item label text content, and then text classification is performed through a proprietary fully connected layer and a CRF model, and the attribute classification results are output: It includes the following two steps: Module M4.1: Automatically extract the features in the word segmentation from the text word vectors with given labels through the Bi-LSTM layer. The Word Embedding sequence of a complete text is used as the input of each time step of the bidirectional LSTM, and then the hidden state sequence output by the forward LSTM is concatenated with the hidden states output by the reverse LSTM at each position to obtain a complete hidden state sequence; Module M4.2: The hidden state is passed into the CRF layer for processing to achieve sequence prediction at the line text level. The CRF layer models the hidden state sequence in the text and calculates the transition probability of each state, and classifies the semantic information of the line text.

Images