Multi-source data-based foreign enterprise legal risk assessment method

Abstract

The invention relates to the technical field of multi-source data processing, in particular to a foreign enterprise legal risk assessment method based on multi-source data, which comprises the following steps of: 1, inputting original data such as texts and pictures related to enterprises to be classified; 2, preprocessing the text data and the picture data respectively; text data usually contains a large amount of irrelevant information and noise, such as punctuation marks, HTML marks, stop words, abbreviation, spelling errors and grammar errors. The information and noise can interfere with the learning and performance of the model, so that the model prediction is inaccurate or inefficient; unnecessary information and noise can be removed through data preprocessing, the invention provides a model for intelligently evaluating the legal risk level, an intelligent algorithm is adopted to replace manual auditing, the time consumed by single evaluation is far less than that consumed by a manual method, meanwhile, the method is suitable for parallel calculation, and large-batch evaluation requirements are supported; the evaluation efficiency is greatly higher than that of a manual evaluation method in the prior art.

Description

Technical Field

[0001] This invention relates to the field of multi-source data processing technology, and in particular to a method for assessing legal risks of foreign-related enterprises based on multi-source data. Background Technology

[0002] With the deepening of global economic integration, foreign trade has become an important way for enterprises to pursue market expansion and competitive advantage. However, the uncertainty and complexity of international trade also lead to many challenges. From goods manufacturing to logistics and transportation, and then to payment settlement, each link can be affected by factors such as market fluctuations, political changes, and natural disasters. These factors may lead to risks such as contract defaults, goods loss, and financial losses, thereby affecting the profitability and reputation of enterprises.

[0003] In traditional foreign trade transactions, risk assessment typically relies on experienced professionals, whose subjective judgment and experience limit the accuracy and consistency of the assessment. Meanwhile, the massive amounts of transaction data and information make it difficult to comprehensively analyze and grasp market dynamics using only human resources. Therefore, innovative methods leveraging artificial intelligence technology have become an important approach to solving these problems. By applying machine learning, data mining, and other technologies to foreign trade risk assessment, hidden correlations can be discovered from big data, improving the accuracy and timeliness of risk prediction.

[0004] In the current technology, the assessment of legal risk levels for foreign-related enterprises can only be determined by staff inspection to determine whether legal risks exist and their levels. This is prone to problems such as incomplete information leading to incorrect risk level assessments, and requires a high level of professional expertise from staff, as well as consuming a large amount of manpower. Summary of the Invention

[0005] The purpose of this invention is to provide a method for assessing legal risks of foreign-related enterprises based on multi-source data. This method employs intelligent algorithms to replace manual review, trains on multi-source features, and allows the model to learn not only image features but also textual information. This enables the model to assess risks from multiple perspectives, including images and text, expanding the information dimensions used in the assessment and significantly reducing the probability of errors in risk level assessment due to incomplete information. More importantly, the time required for a single assessment in this application is far less than that of manual methods. It is also suitable for parallel computing, supporting large-scale assessment needs, and its assessment efficiency is significantly higher than that of existing manual assessment methods.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for assessing legal risks of foreign-related enterprises based on multi-source data, comprising the following steps:

[0007] Step 1: Input the raw data, such as text and images, related to the companies to be categorized;

[0008] Step 2: Perform preprocessing operations on each text and image data separately;

[0009] Text data often contains a lot of irrelevant information and noise, such as punctuation marks, HTML tags, stop words, abbreviations, spelling errors, and grammatical errors. This information and noise can interfere with the model's learning and performance, leading to inaccurate or inefficient predictions. Data preprocessing can remove this unnecessary information and noise, making the text data cleaner and more useful. Image preprocessing methods include, but are not limited to, scaling, translation, transposition, mirroring, rotation, normalization, dimensionality reduction, denoising, equalization, smoothing, and other common methods in the field of image processing, to meet the input conditions of deep learning network models. For example, the input image can be reduced or enlarged to a certain size, such as scaling an original image with a resolution of 1920×1080 to a standard image with a resolution of 224×224.

[0010] Step 3: Encode each preprocessed text and image data separately, converting the information carried by the original data into vector form;

[0011] Step 4: Concatenate the encoded text and image vectors into a whole text feature vector and a whole image feature vector.

[0012] Step 5: Perform a feature fusion of the full-text feature vector and the full-image feature vector using a hybrid attention mechanism to obtain the original fused feature vector t.

[0013] Step 6: Extract features from the original fused feature vector using a convolutional neural network to obtain the high-level fused feature vector;

[0014] Step 7: Perform secondary feature fusion with the high-level fused feature vector, the original full-text feature vector, and the original full-image feature vector;

[0015] Step 8: Input the feature vector after secondary fusion into the classifier to obtain the final enterprise risk level classification result.

[0016] Preferably, step one specifically includes encoding operations on the image training set and the text training set:

[0017] The image training set uses encoding operations (such as...) Figure 1 In the image encoder, traditional convolution operations are used. This application uses a ResNet 50 backbone network for feature extraction. Since convolution feature extraction is local, a self-attention operation is performed after feature extraction by the image backbone network to increase the globality of image features and convert image features into sequence representations that correspond to text sequences.

[0018] Preferably, the text training set is used for encoding operations (such as...). Figure 2 In this application, the Transformer is used for feature extraction (the Text Encoder in the image), which is global and parallel, reducing inference and training time. After encoding, a sequence with the same dimensions as the image sequence is obtained.

[0019] Preferably, step two specifically includes fusing image feature sequences and text feature sequences:

[0020] The specific steps are as follows:

[0021] Determine whether the dimensions of the image sequence and the text feature sequence are consistent;

[0022] Image feature sequences and text feature sequences are normalized to prevent the negative impact of outliers on the model, as shown in formulas a and b:

[0023]

[0024]

[0025] The normalized text sequence and image sequence are multiplied by a dot product, as shown in formula c:

[0026] Feature mix =SEQ img ·SEQ text c.

[0027] Preferably, as shown in formula c, the image feature sequence and the text feature sequence are multiplied by a dot product to obtain a fused feature map. This allows the feature map to contain not only image feature information but also text feature information, enabling the two data sources to work together to predict the legal risk level. Furthermore, during backpropagation, image feature optimization is supervised by text information, and text information optimization is supervised by image information, strengthening the model's extraction of effective features, as shown in formulas d and e.

[0028]

[0029]

[0030] PCA is performed on the fused feature map to reduce the dimensionality of the matrix and extract more effective feature maps.

[0031] The reduced-dimensional feature map is serialized to obtain a sequence with the same dimensions as the image feature sequence and the text feature sequence.

[0032] Preferably, step three includes text and feature relevance supervision;

[0033] In addition to being supervised by staff who have labeled the risk levels, the image-based prediction of legal risk levels is also supervised by the corresponding text features, thus achieving consistency in prediction results across different data sources. The text-based prediction of legal risk levels is consistent with the text itself, as shown in formulas f and g:

[0034]

[0035]

[0036] Preferably, formula f represents calculating the difference between image prediction and text prediction, assuming text prediction is the true legal risk level; formula g represents calculating the difference between text prediction and image prediction, assuming image prediction is the true legal risk level.

[0037] Preferably, step four includes predicting the risk level.

[0038] Based on the sequence dimension, the text prediction sequence, image prediction sequence, and hybrid prediction sequence are fused together, and the final risk level is predicted through MLP.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] 1. This application proposes an intelligent model for assessing the level of legal risk. It uses intelligent algorithms to replace manual review, and the time required for a single assessment is much shorter than that of manual methods. At the same time, it is suitable for parallel computing, supports large-scale assessment needs, and the assessment efficiency is much higher than that of manual assessment methods in the prior art.

[0041] 2. This application proposes a method for assessing the legal risk level of foreign-related enterprises based on multi-source data. This method allows the model to learn not only image features but also textual information, enabling the model to assess from multiple aspects such as images and text. This expands the information dimensions used in the assessment and greatly reduces the probability of errors in risk level assessment caused by incomplete information.

[0042] 3. This application uses textual and image information as supervision to train the model, which improves the accuracy of risk level assessment based on single-source data in text form commonly used in the prior art. Attached Figure Description

[0043] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0044] Figure 1This is a data processing flowchart of the present invention;

[0045] Figure 2 This is a diagram of the legal risk level prediction network architecture of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Please see Figures 1 to 2 The present invention provides a technical solution:

[0048] A method for assessing legal risks for foreign-related enterprises based on multi-source data includes the following steps:

[0049] Step 1: Input the raw data, such as text and images, related to the companies to be categorized;

[0050] Step 2: Perform preprocessing operations on each text and image data separately;

[0051] Text data often contains a lot of irrelevant information and noise, such as punctuation marks, HTML tags, stop words, abbreviations, spelling errors, and grammatical errors. This information and noise can interfere with the model's learning and performance, leading to inaccurate or inefficient predictions. Data preprocessing can remove this unnecessary information and noise, making the text data cleaner and more useful. Image preprocessing methods include, but are not limited to, scaling, translation, transposition, mirroring, rotation, normalization, dimensionality reduction, denoising, equalization, smoothing, and other common methods in the field of image processing, to meet the input conditions of deep learning network models. For example, the input image can be reduced or enlarged to a certain size, such as scaling an original image with a resolution of 1920×1080 to a standard image with a resolution of 224×224.

[0052] Step 3: Encode each preprocessed text and image data separately, converting the information carried by the original data into vector form;

[0053] Step 4: Concatenate the encoded text and image vectors into a whole text feature vector and a whole image feature vector.

[0054] Step 5: Perform a feature fusion of the full-text feature vector and the full-image feature vector using a hybrid attention mechanism to obtain the original fused feature vector t.

[0055] Step 6: Extract features from the original fused feature vector using a convolutional neural network to obtain the high-level fused feature vector;

[0056] Step 7: Perform secondary feature fusion with the high-level fused feature vector, the original full-text feature vector, and the original full-image feature vector;

[0057] Step 8: Input the feature vector after secondary fusion into the classifier to obtain the final enterprise risk level classification result.

[0058] Step one specifically includes encoding operations on the image training set and the text training set:

[0059] The image training set uses encoding operations (such as...) Figure 1 In the image encoder, traditional convolution operations are used. This application uses a ResNet 50 backbone network for feature extraction. Since convolution feature extraction is local, a self-attention operation is performed after feature extraction by the image backbone network to increase the globality of image features and convert image features into sequence representations that correspond to text sequences.

[0060] Encoding operations on the text training set (e.g.) Figure 2 In this application, the Transformer is used for feature extraction (the Text Encoder in the image), which is global and parallel, reducing inference and training time. After encoding, a sequence with the same dimensions as the image sequence is obtained.

[0061] Step two specifically involves fusing image feature sequences and text feature sequences:

[0062] The specific steps are as follows:

[0063] Determine whether the dimensions of the image sequence and the text feature sequence are consistent;

[0064] Image feature sequences and text feature sequences are normalized to prevent the negative impact of outliers on the model, as shown in formulas a and b:

[0065]

[0066]

[0067] The normalized text sequence and image sequence are multiplied by a dot product, as shown in formula c:

[0068] Feature mix =SEQ img ·SEQ text c.

[0069] As shown in formula c, the image feature sequence and the text feature sequence are multiplied by a dot product to obtain a fused feature map. This feature map contains not only image feature information but also text feature information, allowing the two data sources to work together to predict the legal risk level. Furthermore, during backpropagation, image feature optimization is supervised by text information, and text information optimization is supervised by image information, strengthening the model's extraction of effective features, as shown in formulas d and e.

[0070]

[0071]

[0072] PCA is performed on the fused feature map to reduce the dimensionality of the matrix and extract more effective feature maps.

[0073] The reduced-dimensional feature map is serialized to obtain a sequence with the same dimensions as the image feature sequence and the text feature sequence.

[0074] Step three includes text and feature relevance supervision;

[0075] In addition to being supervised by staff who have labeled the risk levels, the image-based prediction of legal risk levels is also supervised by the corresponding text features, thus achieving consistency in prediction results across different data sources. The text-based prediction of legal risk levels is consistent with the text itself, as shown in formulas f and g:

[0076]

[0077]

[0078] Formula f represents calculating the difference between image prediction and text prediction, assuming text prediction is the true legal risk level; formula g represents calculating the difference between text prediction and image prediction, assuming image prediction is the true legal risk level.

[0079] Step four includes predicting the risk level:

[0080] Based on the sequence dimension, the text prediction sequence, image prediction sequence, and hybrid prediction sequence are fused together, and the final risk level is predicted through MLP.

[0081] This application realizes automatic legal risk level prediction based on multi-source features (text and image features are used in this application). The network architecture diagram of the legal risk level prediction of this application is as follows. Figure 2As shown, firstly, the input image and text are encoded and serialized separately; then, the text serialization vector and the image serialization vector are fused to obtain a new feature map; next, the feature map containing the information of text and image serialization is scaled to obtain a sequence of the same dimension as the image and text; the image, text, and the fused sequence are then superimposed by channel to obtain a new sequence; finally, the prediction result is obtained through the MLP method.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for assessing legal risks of foreign-related enterprises based on multi-source data, characterized in that: Includes the following steps: Step 1: Input the raw data, such as text and images, related to the companies to be categorized; Step 2: Perform preprocessing operations on each text and image data separately; Text data often contains a lot of irrelevant information and noise, such as punctuation marks, HTML tags, stop words, abbreviations, spelling errors, and grammatical errors; This information and noise can interfere with the model's learning and performance, leading to inaccurate or inefficient predictions. Data preprocessing can remove this unnecessary information and noise, making the text data cleaner and more useful. Image preprocessing methods include, but are not limited to, scaling, translation, transposition, mirroring, rotation, normalization, dimensionality reduction, denoising, equalization, smoothing, and other common methods in the field of image processing, to meet the input conditions of deep learning network models. For example, the input image can be reduced or enlarged to a certain size, such as scaling the original image with a resolution of 1920×1080 to a standard image with a resolution of 224×224. Step 3: Encode each preprocessed text and image data separately, converting the information carried by the original data into vector form; Step 4: Concatenate the encoded text and image vectors into a whole text feature vector and a whole image feature vector; Step 5: Perform a feature fusion of the full text feature vector and the full image feature vector using a hybrid attention mechanism to obtain the original fused feature vector t; Step 6: Extract features from the original fused feature vector using a convolutional neural network to obtain the high-level fused feature vector; Step 7: Perform secondary feature fusion with the high-level fused feature vector, the original full-text feature vector, and the original full-image feature vector; Step 8: Input the feature vector after secondary fusion into the classifier to obtain the final enterprise risk level classification result.

2. The method for assessing legal risks of foreign-related enterprises based on multi-source data as described in claim 1, characterized in that: Step one specifically includes encoding operations on the image training set and the text training set: The image training set employs encoding operations, traditional convolution operations, and a ResNet 50 backbone network for feature extraction. Since convolutional feature extraction is localized, a self-attention operation is performed after feature extraction by the image backbone network to increase the globality of image features, transforming image features into sequence representations that correspond to text sequences.

3. The method for assessing legal risks of foreign-related enterprises based on multi-source data as described in claim 2, characterized in that: The text training set is encoded, and the Transformer is used for feature extraction. This process is global and parallel, reducing inference and training time. The encoding process yields a sequence with the same dimensions as the image sequence.

4. The method for assessing legal risks of foreign-related enterprises based on multi-source data as described in claim 3, characterized in that: Step two specifically involves fusing image feature sequences and text feature sequences: The specific steps are as follows: Determine whether the dimensions of the image sequence and the text feature sequence are consistent; Image feature sequences and text feature sequences are normalized to prevent the negative impact of outliers on the model, as shown in formulas a and b: The normalized text sequence and image sequence are multiplied by a dot product, as shown in formula c: Feature mix =SEQ img ·SEQ text c。 5. The method for assessing legal risks of foreign-related enterprises based on multi-source data as described in claim 4, characterized in that: As shown in formula c, the image feature sequence and the text feature sequence are multiplied by a dot product to obtain a fused feature map. This feature map contains not only image feature information but also text feature information. The two data sources work together to predict the legal risk level. Furthermore, during backpropagation, image feature optimization is supervised by text information, and text information optimization is supervised by image information, strengthening the model's extraction of effective features, as shown in formulas d and e. PCA is performed on the fused feature map to reduce the dimensionality of the matrix and extract more effective feature maps. The reduced-dimensional feature map is serialized to obtain a sequence with the same dimensions as the image feature sequence and the text feature sequence.

6. The method for assessing legal risks of foreign-related enterprises based on multi-source data as described in claim 5, characterized in that: Step three includes text and feature relevance supervision; In addition to being supervised by staff who have labeled the risk levels, the image-based prediction of legal risk levels is also supervised by the corresponding text features, thus achieving consistency in prediction results across different data sources. The text-based prediction of legal risk levels is consistent with the text itself, as shown in formulas f and g:

7. The method for assessing legal risks of foreign-related enterprises based on multi-source data as described in claim 6, characterized in that: Formula f represents calculating the difference between image prediction and text prediction, assuming text prediction is the true legal risk level; formula g represents calculating the difference between text prediction and image prediction, assuming image prediction is the true legal risk level.

8. The method for assessing legal risks of foreign-related enterprises based on multi-source data according to claim 7, characterized in that: Step four includes predicting the risk level: Based on the sequence dimension, the text prediction sequence, image prediction sequence, and hybrid prediction sequence are fused together, and the final risk level is predicted through MLP.

Images