KR102574459B1 - Device and method for electronic document management based on artificial intelligence having automatic notification function - Google Patents

Abstract

One object of the present invention is to provide an artificial intelligence-based electronic document management device that is applied to services such as contract management and report management to manage and analyze electronic documents and provide automatic notification of specific events.
To this end, an artificial intelligence-based electronic document management device including an automatic notification function according to an embodiment of the present invention receives an electronic document image and matches text included in the electronic document image for each item to obtain information about a plurality of items. one or more artificial neural network modules for outputting; a selection module for selecting item information related to a date from among the plurality of item information; and an automatic notification item determination module for determining automatic notification target item information to be automatically notified among the selected item information.

Description

Device and method for electronic document management based on artificial intelligence having automatic notification function

The present invention relates to an artificial intelligence-based electronic document management apparatus and method including an automatic notification function.

Due to the development of the cloud market, electronic document businesses such as Docusign and Modusign have recently started to flourish, and various problems of existing paper documents are being solved by the electronic document business. For example, in Korea, paper documents generated at one bank branch reach 100 million annually, requiring 73,000 boxes and a very large storage space of 600 pyeong. Compared to paper documents, which take up a very large space, electronic documents You can manage more documents in a lot less space. In addition, when electronic documents are used, management costs due to document production, distribution, storage, disposal, etc., as well as printing and ink costs, can be reduced, and document management and information leakage prevention are made easier.

In the process of transitioning from the paper document business to the electronic document business, the important point is to digitize the existing paper documents. The technology used at this time is Optical Character Recognition (OCR) technology. OCR is a technology that recognizes and extracts text from an image through scanning. OCR was actively applied in the financial sector in the early 2000s. At that time, banks were mostly responsible for processing paper documents printed on paper slips or manually entered by customers, and it was an important process to increase the recognition rate by training the OCR engine on documents provided by customers. Parameters had to be adjusted, and a lot of resources were spent on correcting misrecognized parts, resulting in a decrease in overall work accuracy.

Since then, the electronic document management business has begun to open a new horizon with the development of artificial intelligence and blockchain. In particular, in relation to artificial intelligence, CNN (Convolutional Neural Network), which was first presented by Y.LeCun in 1989, started with Alex Khrizhebsky's AlexNet in 2012, and the AI era began in earnest. Recently, VGG-16, ResNet, Various deep learning models (or deep models) such as EfficientNet, AmoebaNet, SqueezeNet, MobileNet, ShuffleNet, MNasNet, FBNet, CharmNet, DenseNet, and Xception have begun to be presented. Since then, the electronic document management business has started to utilize this deep model, which has had the effect of overcoming the limitations of the existing OCR technology.

Republic of Korea Patent Publication No. 10-2021-0086849, method for generating documents, Legal Insight Co., Ltd. Korean Registered Patent No. 10-2289935, artificial intelligence-based legal document analysis system and method, Intellicon Research Institute

Accordingly, an object of the present invention is to provide an artificial intelligence-based electronic document management device and method for managing and analyzing electronic documents applied to services such as contract management and report management.

Hereinafter, specific means for achieving the object of the present invention will be described.

An artificial intelligence-based electronic document management device including an automatic notification function according to an embodiment of the present invention receives an electronic document image and outputs a plurality of item information obtained by item-by-item matching of text included in the electronic document image. one or more artificial neural network modules for; a selection module for selecting item information related to a date from among the plurality of item information; and an automatic notification item determination module for determining automatic notification target item information to be automatically notified among the selected item information.

In the artificial intelligence-based electronic document management device described above, the automatic notification item determination module includes a transformer-based artificial neural network module that receives the selected item information and outputs item information for automatic notification, The transformer-based artificial neural network module receives future expected compensation, state information, and past outputs of the artificial neural network module, and generates an optimal output including automatic notification target item information as an output, and the future expected compensation determines the automatic notification item. The module represents the total amount of rewards expected in the future by generating the automatic notification target item information, the status information includes at least the selected item information received by the automatic notification item determination module, and the past output is the transformer It includes the automatic notification target item information output in the past by the base artificial neural network module.

In the aforementioned artificial intelligence-based electronic document management device, the status information includes legal information associated with the electronic document image.

In the artificial intelligence-based electronic document management apparatus described above, the transformer-based artificial neural network module includes a first block and a second block, and the first block receives the future expected compensation and the state information and receives final encoding information. is output and provided as an input to the second block, and the second block receives the past output and generates an optimal output including the automatic notification target item information.

In the aforementioned artificial intelligence-based electronic document management device, the at least one artificial neural network module includes an artificial neural network that receives an electronic document image, extracts text included in the electronic document image, and outputs text information. module; and a text processing module including an artificial neural network that receives entire text information, which is information in which the text information is sequentially integrated in the order of y-coordinates, and outputs item information obtained by item-by-item matching of the text included in the electronic document image. includes

As described above, the present invention has the following effects.

First, according to an embodiment of the present invention, an effect of learning an artificial neural network for electronic document management without leakage of personal information is generated.

Second, according to an embodiment of the present invention, an effect of being able to classify electronic document items into customized classes according to the types of electronic documents occurs.

Thirdly, according to one embodiment of the present invention, an automatic notification can be performed by selecting item information requiring automatic notification among item information of an electronic document.

The following drawings attached to this specification illustrate preferred embodiments of the present invention, and together with the detailed description of the invention serve to further understand the technical idea of the present invention, the present invention is limited only to those described in the drawings. and should not be interpreted.
1 is a schematic diagram showing the operational relationship of an artificial intelligence-based electronic document management device according to an embodiment of the present invention.
2 is a schematic diagram showing an operating relationship according to each step of an artificial intelligence-based electronic document management device and an electronic document management server according to an embodiment of the present invention.
3 is a schematic diagram showing the configuration of an artificial intelligence-based electronic document management device according to an embodiment of the present invention.
4 is a schematic diagram showing a specific operational relationship of an artificial intelligence-based electronic document management device according to an embodiment of the present invention.
5 is a schematic diagram showing a specific configuration of a text extraction module according to an embodiment of the present invention.
6 is a schematic diagram showing a text box generating neural network module according to an embodiment of the present invention.
7 is a schematic diagram showing a text box segmentation module according to an embodiment of the present invention.
8 is a schematic diagram showing an operating relationship of a segmental vector generation module according to an embodiment of the present invention.
9 is a schematic diagram showing an operating relationship of a segmental vector generation module according to a modified example of the present invention.
10 is a schematic diagram showing a text sequence generating neural network module according to an embodiment of the present invention.
11 is a schematic diagram showing a reinforcement learning module according to an embodiment of the present invention.
12 is a schematic diagram showing a reinforcement learning module according to a modified example of the present invention.
13 is a flowchart illustrating an operation example of a reinforcement learning module according to an embodiment of the present invention.
14 is a schematic diagram illustrating a text processing module according to an embodiment of the present invention.
15 is a schematic diagram showing an electronic document classification neural network module according to an embodiment of the present invention.
16 is a schematic diagram illustrating an item classification neural network module according to an embodiment of the present invention.
17 is a schematic diagram showing a neural network processing module according to an embodiment of the present invention.
18 is a schematic diagram illustrating a correction information generating module according to an embodiment of the present invention.
19 is a schematic diagram showing an abnormal text classification process of an abnormal text classification module according to an embodiment of the present invention.
20 is a schematic diagram showing the structure of a normal text generating module according to an embodiment of the present invention.
21 schematically illustrates an artificial intelligence-based electronic document management device having an automatic notification function according to an embodiment of the present invention.
22 schematically illustrates an automatic notification module according to an embodiment of the present invention.
23 is a diagram schematically illustrating a transformer-based artificial neural network module included in an automatic notification item determination module according to an embodiment of the present invention.

Hereinafter, an embodiment in which a person skilled in the art can easily practice the present invention will be described in detail with reference to the accompanying drawings. However, in the detailed description of the operating principle of the preferred embodiment of the present invention, if it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description will be omitted.

In addition, the same reference numerals are used for parts having similar functions and actions throughout the drawings. Throughout the specification, when a specific part is said to be connected to another part, this includes not only the case where it is directly connected but also the case where it is indirectly connected with another element interposed therebetween. In addition, including a specific component does not exclude other components unless otherwise stated, but means that other components may be further included.

In the following description of the invention, artificial neural networks may be described in terms such as artificial neural network modules, neural network modules, deep learning, deep learning models, deep models, and models.

In the following description of the invention, the convolutional neural network, which is a neural network using convolutional products, may be described as CNN, ConvNet, and the like.

In the following description of the invention, electronic documents may refer to electronic documents such as contracts, estimates, invoices, and reports.

AI-based electronic document management device

Regarding the operating relationship of the artificial intelligence-based electronic document management device according to an embodiment of the present invention, Figure 1 is a schematic diagram showing the operating relationship of the artificial intelligence-based electronic document management device according to an embodiment of the present invention, 2 is a schematic diagram showing an operating relationship according to each step of an artificial intelligence-based electronic document management device and an electronic document management server according to an embodiment of the present invention, and FIG. 3 is an artificial intelligence-based document management server according to an embodiment of the present invention. Figure 4 is a schematic diagram showing the configuration relationship of the electronic document management device of the present invention, Figure 4 is a schematic diagram showing the specific operating relationship of the artificial intelligence-based electronic document management device according to an embodiment of the present invention.

As shown in FIG. 1, an electronic document image of a specific document (eg, a real estate sales contract, a real estate lease agreement, etc.) is generated through a photograph of the user client 100 or a scan of a separate scanning device, so that the user client ( 100) is input to the artificial intelligence-based electronic document management device 1 according to an embodiment of the present invention, and the artificial intelligence is sent to the electronic document management server 200 from the artificial intelligence-based electronic document management device 1. It is configured to transmit parameters of an artificial neural network included in the based electronic document management device 1.

In addition, as shown in FIG. 2, the artificial intelligence-based electronic document management apparatus 1 according to an embodiment of the present invention is configured in each of a plurality of user clients 100, and the electronic document management server 200 and the wired/wireless It may be connected to a network and configured to perform a client learning step, a parameter update step, and a main neural network download step.

In addition, as shown in FIGS. 3 and 4 , the artificial intelligence-based electronic document management device 1 according to an embodiment of the present invention may be configured to include a text extraction module, a text processing module, and a neural network processing module. And, the electronic document management server 200 may be configured to include a main neural network module 210 and a federated learning module 220. According to an example, the neural network processing module of the artificial intelligence-based electronic document management device 1 transmits the changed parameters and modified change information to the federation learning module 220 of the electronic document management server 200, and the federated learning module ( In step 220, a pretrained main neural network may be generated by performing federated learning based on parameters and modified change information received from the plurality of user clients 100. The pretrained main neural network may be configured to update the text extraction module and the artificial neural network module of the text processing module through the main neural network module 210 . According to this, an effect of being able to train the artificial neural network of the text extraction module and the text processing module for performing electronic document management without sharing personal information included in the electronic document with other user clients occurs.

Hereinafter, the operating principle of the text extraction module of the artificial intelligence-based electronic document management device 1 will be described in more detail. 5 is a schematic diagram showing a specific configuration of a text extraction module according to an embodiment of the present invention.

As shown in FIG. 5 , the text extraction module is a module that receives an electronic document image, extracts text included in the electronic document image, and outputs text information. A text extraction module according to an embodiment of the present invention may include a text box generation neural network module, a text box segmentation module, a text sequence generation neural network module, and a text information generation module.

6 is a schematic diagram showing a text box generating neural network module in a text extraction module according to an embodiment of the present invention.

As shown in FIG. 6, the text box generation neural network module uses an electronic document image as input data and a text class that estimates whether text is included in a text box, which is a bounding box surrounding the text included in the electronic document image. Text box information including text class information meaning probability, text coordinate information including the coordinates, height, width and angle of the text box, and text size class information meaning the estimated font size of the text as output data It may mean an artificial neural network module.

The text box generating neural network module according to an embodiment of the present invention may include a multi-object detection artificial neural network module that performs classification and localization, and this multi-object detection artificial neural network module includes RCNN as a 2-stage detector ( 2013), OverFeat (ICLR 2014), Fast RCNN (ICCV 2015), Faster RCNN (NIPS 2015), Mask RCNN (ICCV 2017), etc. can be used, and as a 1-stage detector, anchor-based YOLO v1 (CVPR 2016) , YOLO v2 (CVPR 2017), YOLO v3 (arXiv 2018), SSD (ECCV 2016), RetinaNet (ICCV 2017), etc. can be used, and as a 1-stage detector, non-anchor based CornerNet (ECCV 2018), ExtreamNet (2019), CenterNet (2019), etc. can be utilized, and CRAFT (down sampling/up sampling) can be utilized.

Output data of the text box generating neural network module according to an embodiment of the present invention may be one or more text box images for at least one object in an electronic document image, and such text box images may include at least one or more text box images included in the electronic document image. It may include text class information (confidence score or class probability), which means the confidence of the text class for an object, and text coordinate data, which is the coordinate information of the corresponding text object (object inferred as a text class). and the text coordinate information of the object includes the coordinates of the top-left corner and bottom-right corner of the bounding box, the coordinates of the centeral region of the bounding box, the width and height of the bounding box, and the angle of the bounding box according to the configuration of the artificial neural network. It can be configured to include text size class information that means the estimated character size of text within the bounding box. For example, text class information for a specific object included in an electronic document image is [0.8], text coordinate information is [x, y, w, h, θ], text size class information is [3], etc. can be configured. According to an embodiment of the present invention, since 'angle' information is further included in the text coordinate information, the text angle of the text box image is standardized.

6 is a schematic diagram showing the structure of a text box generating neural network module according to an embodiment of the present invention composed of YOLO v1 (CVPR 2016). For example, as shown in FIG. 6, when the text box generating neural network module according to one embodiment of the present invention is configured with YOLO v1 (CVPR 2016), DarkNet Architecture is used, and feature maps are generated through convolution layers. It is configured to infer text class probability (class confidence, text class information), coordinate data (text coordinate information) and text size class information of the bounding box directly through a fully connected layer and output them as output data. In YOLO, the electronic document image, which is the input image, is divided into SxS grids, and the bounding boxes (SxSxB) corresponding to each grid area, Class confidence (Probability (object) × IoU (prediction, ground truth)), Class probability map (Probability (Class_i |object)). For example, a specific network structure can be configured to output 5 bounding box coordinates (text coordinate information) and confidence score (text class information) per grid area, and an electronic document image can be configured to be input in a size of 448x448x3. Activation map of DarkNet Architecture can be configured in size of 7x7x1024, and Fully Connected Layers of 4096 and 7x7x30 can be configured after DarkNet Architecture.

7 is a schematic diagram showing a text box segmentation module in a text extraction module according to an embodiment of the present invention.

As shown in FIG. 7, the text box segmentation module takes as input data a text box image, which is a part of an electronic document image corresponding to text box information, output by the text box generation neural network module in the example of FIG. 6, A CNN module including a fully connected layer (FC layer) or preferably a fully convolutional layer (FCN layer) that uses the feature map (activation map, activation information) for the text box image as output data, and an integration that integrates the activation information Segment vector generation that generates a text box feature vector, which is activation information, as an output, and generates n segment vector columns (segment vector information) formed by segmenting the text box feature vector into n vectors (a specific number of segments) in the longitudinal direction. modules may be included.

8 is a schematic diagram showing an operating relationship of a segmental vector generation module according to an embodiment of the present invention.

As shown in FIG. 8, the segmentation vector generation module is connected to a plurality of Conv.layers configured in the CNN module including the FCN layer, and when the input data of the CNN module including the FCN layer is a text box image of a specific size, Conv. . A plurality of activation information (including activation values for each coordinate) including at least one-dimensional activation map output from the layer is input, coordinate data (text coordinate information) of the text class for the text box image is input, and multiple The size of the activation information is adjusted and integrated to correspond to the text coordinate information to generate a text box feature vector, which is integrated activation information, and segmented between the peaks of the activation values in the length direction of the text box to create n segment vector columns (segmental vector columns). vector information). At this time, the CNN module including the FCN layer is configured to be pre-learned in a state in which an output layer that outputs the text class probability of the text box image by applying Global Average Pooling (GAP) to the FCN layer is configured. In addition, the segment vector generation module according to an embodiment of the present invention expands the size (w, h) of the text box image, which is input data, by a specific weight according to the text class probability output from the CNN module including the FCN layer. It can be configured to feed back with input data. In this case, for example, the specific weight may be configured in the form of 1/[text class probabilty] or 1+(1-[text class probabilty]). For example, if the text class probability output by the text box image input to the CNN module including the FCN layer is 0.7, the text box image may be expanded by a weight of 1.3 and inputted as input data.

According to the text box segmentation module according to an embodiment of the present invention, each character included in the text box image can be precisely inferred as a separate segment vector column. In addition, since the FCN layer and the GAP are configured, even if the size of the text box image, which is input data, is configured in various ways, a text box feature vector can be generated. In addition, since the text class probability is inferred in the text box generation neural network module and the text box segmentation module, respectively, the CNN module including the FCN layer of the text box segmentation module is derived based on the text class probability previously inferred in the text box generation neural network module. There is an effect that can be learned.

9 is a schematic diagram showing an operating relationship of a segmental vector generation module according to a modified example of the present invention.

As shown in FIG. 9, according to a modified example of the present invention, a CNN module including an FCN layer is composed of an artificial neural network composed of a single convolutional network that outputs text class probability as output data through global average pooling, It further includes a class activation generation module that reduces dimensions between the segment vector generation module and the CNN module including the FCN layer, and the class activation generation module is connected to a plurality of conv.layers of the CNN module including the FCN layer to connect the FCN layer. When the input data of the CNN module including the layer is a text box image, the activation information (including the activation value for each coordinate) including the activation map of at least one dimension output from the Conv.layer is input and the activation map corresponding to the text class is received. It can be learned with a CNN module that includes an FCN layer configured to output class activation information (including activation values for each coordinate according to text class) and output text class probability through a global average pooling function. In addition, the class activation information and text coordinate information are input to the segment vector generation module, and the segment vector generation module adjusts and integrates the size of the input class activation information to correspond to the size of the text coordinate information and integrates the text box as integrated activation information. A feature vector may be generated, and n segment vector columns (segment vector information) may be generated by segmenting between peaks of activation values in the longitudinal direction of the text box.

At this time, generation of class activation information by the class activation generation module may be performed as follows.

In Equation 1 above, M _c (x, y) is the activation value located at (x, y) that affects classification into class c, w _k ^c is the weight of the k-th channel for class c in the activation map, f _k (x, y) means an activation value located at (x, y) of the k-th channel of the activation map.

According to the modified example of the present invention, since the text box feature vector limited to the text class is output using the class activaiton map generated limited to the text class instead of the activation map for the entire class, the character included in the text box image The effect of being able to more precisely infer each one as a separate segment vector column is generated. In addition, since a single artificial neural network module can be configured from a CNN module including an FCN layer to a class activation generation module, the efficiency of artificial neural network learning and inference is improved. In addition, since the FCN layer and the GAP are configured, even if the size of the text box image, which is input data, is configured in various ways, a text box feature vector can be generated. In addition, since the text class probability is inferred in the text box generation neural network module and the text box segmentation module, respectively, the CNN module including the FCN layer of the text box segmentation module is derived based on the text class probability previously inferred in the text box generation neural network module. There is an effect that can be learned.

The text sequence generation neural network module in the text extraction module uses the sequential segment vector column (vector column segmented into one syllable) for the text box image as input data and the m-th segment text information (vector segmented into one syllable). It may refer to a recursive artificial neural network module that outputs text sequence information sequentially including information about segmented text, which is text about text, as output data. At this time, the recurrent artificial neural network module may be configured with an artificial neural network structure such as a recurrent neural network (RNN), long-short term memory (LSTM), and Bi-LSTM.

10 is a schematic diagram showing a text sequence generating neural network module according to an embodiment of the present invention.

As shown in FIG. 10, the text sequence generating neural network module of the text extraction module according to an embodiment of the present invention is composed of a plurality of RNN blocks, and one RNN block includes a first RNN cell and a second RNN cell. can do. As shown in FIG. 10, the first RNN cell receives the m-th segment vector column, the output data of the previous cell, and the text box feature vector as input data, and receives language type information for the m-th segment vector column (for example, Korean (kor) / English (eng) / Japanese (jp) / number (num), etc.) are output, and the second RNN cell displays the language type information for the corresponding m-th segment vector column and the hidden state of the first RNN cell. It may be configured to receive input and output m-th segment text information. The text sequence generation neural network module according to an embodiment of the present invention performs inference when the language type information output from the first RNN cell is 'eng' and when the rewards for all actions calculated by the reinforcement learning module are negative numbers. can be terminated According to the text sequence generation neural network module according to an embodiment of the present invention, it is configured to sequentially output language type information and m-th segment text information according to the order of character description (from left to right), and the language type generated in the previous step. Since the information and the segmented text information affect the segmented text information to be generated in the next step, the effect of generating the segmented text more accurately occurs. In addition, since the text box feature vector is input in the first RNN cell and segmented text information for one syllable is inferred in the second RNN cell based on this, segmented text in the overall context of the sentence in the text box can be inferred. The effect of being able to happen.

The RNN block of the text sequence generating neural network module according to an embodiment of the present invention may be learned by the reinforcement learning module. The reinforcement learning module that learns the RNN block sets the pre-generated text box feature vector, language type vector, m-1 th segment vector column, and m-1 th segment text information as an environment, and the text sequence generation neural network module Each RNN block is an agent, and the situation when the language type information and segment text information from the 1st segment vector column to the n-1 th segment vector column is output is the state. The language type information and the m-th segment text information output for the th segment vector column are set as Actions, and the similarity between the language type information and the m-th segment text information of the current step, which are the output data of the current step, and the ground truth (for example, The higher the cosine similarity or the smaller the difference (eg, Kullback-Leibler divergence), the higher the reward is generated, which can be configured to update the hidden layer of the RNN block as an agent.

11 is a schematic diagram showing a reinforcement learning module according to an embodiment of the present invention.

As shown in FIG. 11, the pre-generated text box feature vector, language type vector, m-1 th segment vector column, and m-1 th segment text information are set as Environment, and each RNN block of the text sequence generating neural network module is the Agent, and the situation when the language type information and segment text information from the 1st segment vector column to the n-1st segment vector column is output is the State. The language type information and the m-th segment text information output for the column are set as Actions, and the language type information and the m-th segment text information of the current step, which are the output data of the current step, and the similarity between the ground truth (eg, cosine similarity) It can be configured to update the hidden layer of the RNN block, which is an agent, by generating a higher reward as the difference is higher or the difference (eg, Kullback-Leibler divergence) is smaller. An RNN block optimized by the reinforcement learning module according to an embodiment of the present invention may be configured such that the hidden layer is fixed.

According to this, an effect of generating optimal segmentation text information corresponding to the feature vector of the image text box according to the sequence of each character in the text box is generated by the text sequence generating neural network module and the reinforcement learning module. In addition, since the reinforcement learning module is configured to be sequentially optimized for each character of the text box without considering all the number of cases for all language type information and m-th segment text information that can be generated by the text sequence generation neural network module, reinforcement The number of cases to be calculated by the learning module is reduced, resulting in a reduction in computing resources.

The reinforcement learning module according to the modified example of the present invention can be configured so that the RNN block is updated by more effective reinforcement learning through the following configuration. 12 is a schematic diagram showing a reinforcement learning module according to a modified example of the present invention.

As shown in FIG. 12, the reinforcement learning module according to the modified example of the present invention is an artificial neural network that learns a value function outputting a value in a specific state, language type information, and m-th segment text information. It may include a policy network that learns a policy function that outputs each probability of , and the policy network and value network according to the modified example of the present invention may be configured to be connected to a specific RNN block of a text sequence generating neural network module. The policy network and the value network can be connected to the RNN block to output language type information and segmental text information for a specific sequence.

The policy network is an artificial neural network that determines the probability of selected language type information and m-th segment text information in each state of the reinforcement learning module. output probabilities. The cost function of the policy network may be a function obtained by calculating cross entropy by multiplying the policy function by the cost function of the value network and then taking the policy gradient. For example, it may be configured as in Equation 2 below. The policy network can be back propagated based on the product of the cross entropy and the time difference error, which is the cost function of the value network.

In Equation 2, π is a policy function, θ is a policy network parameter, π _θ (a _i │s _i ) is the probability of taking a specific action (language type information and m-th segment text information) in the current episode, and V is a value function , w is a value network parameter, s _i is the state information of the current episode i, S _i+1 is the state information of the next episode i+1, r _i+1 is the reward expected to be obtained in the next episode, V _w (s _i ) may mean the possibility of compensation in the current episode, V _w (s _i+1 ) may mean the possibility of compensation in the next episode, and γ may mean the depreciation rate. In this case, r _i+1 may be configured to receive the language type information of the current step and the similarity between the m-th segment text information and the ground truth.

In the policy network according to an embodiment of the present invention, prior language type information and m-th segment text information and result information (the language type information of the current step and the m-th segment text information and the ground truth) are used before reinforcement learning proceeds. The basis of the policy can be learned by updating the weight of the policy network through supervised learning based on similarity). That is, the weight of the policy network can be set through supervised learning based on language type information, m-th segment text information, and performance information. According to this, the policy network can be learned very quickly by the history of the language type information and the m-th segment text information.

In addition, according to an embodiment of the present invention, during supervised learning of a policy network, supervised learning may be performed based on type information and parameter information of an operation unit of a previous layer, including random vectors, and result information accordingly. The random vector may use, for example, a Gaussian distribution. According to this, the policy network can output challenging language type information and m-th segment text information with random probability. When supervised learning of the policy network is configured so that supervised learning is based on previous language type information, m-th segment text information, and result information, the selection of language type information and m-th segment text information is optimized within the policy of the previous step. results will appear. However, according to an embodiment of the present invention, if a random vector is included during supervised learning of the policy network, as reinforcement learning progresses, the policy network can learn more effective language type information and m-th segment text information than the policy in the previous step. The effect of being able to happen.

The value network is an artificial neural network that derives the possibility of achieving a reward in each state that the reinforcement learning module can have, and learns a value function. The value network presents the direction in which the RNN block, which is the agent, will be updated. To this end, the input variable of the value network is set to state information, which is information about the state of the reinforcement learning module, and the output variable of the value network is set to reward possibility information, which is the possibility of the RNN block achieving reward (language type information of the current step and similarity between the m-th segment text information and the ground truth). Compensation possibility information according to an embodiment of the present invention may be calculated by a Q-function as shown in the following equation.

In Equation 3 above, Q _π means information on the possibility of compensation expected in the future in the case of state s and action a in a specific policy π, R may mean compensation for a specific period, and gamma may mean depreciation rate. S _t may mean a state at time t, A _t may mean an action at time t, and E may mean an expected value. Compensation possibility information (Q value) according to an embodiment of the present invention defines the update direction and size of the policy network.

At this time, the cost function of the value network may be a mean square error (MSE) function for the value function, and may be configured as shown in Equation 4 below, for example. The value network can be back propagated based on the time lag error, which is the cost function of the value network.

In Equation 4, V is a value function, w is a value network parameter, s _i is the state information of the current episode i, S _{i + 1} is the state information of the next episode i + 1, r _{i + 1} is the next episode A reward expected to be obtained, V _w (s _i ) may mean a reward possibility in the current episode, V _w (s _{i + 1} ) may mean a reward possibility in the next episode, and γ may mean a depreciation rate. In this case, r _i+1 may be configured to receive the language type information of the current step and the similarity between the m-th segment text information and the ground truth.

Accordingly, the value network can update the cost function of Equation 3 in the direction of gradient descent when the state of the reinforcement learning module is changed.

According to an embodiment of the present invention, while learning the value network separately from the policy network, the Q value of the value network is supervised instead of starting from random, so that rapid learning is possible. According to this, an effect of greatly reducing the burden of exploration occurs in an action of selecting a combination of very complex language type information and m-th segment text information.

According to the reinforcement learning module according to an embodiment of the present invention, when the policy network that has completed supervised learning selects the language type information and the m-th segment text information of the current episode i, the value network selects the selected language type information and the m-th segment text information. It is trained to predict the reward (the language type information of the current step and the similarity between the m-th segment text information and the ground truth) when processing text information. The policy network and value network of the reinforced learning module that has been trained are combined with simulation using RNN blocks, and are finally used to select language type information and m-th segment text information.

In addition, according to the value network according to an embodiment of the present invention, the update of the policy network outputting the probability of the selected language type information and the m-th segment text information can be performed every episode. In the existing reinforcement learning, there is a problem in that the reinforcement learning model is updated after all episodes are finished, so it is difficult to apply the language type information and the m-th segment text information to the RNN module that sequentially generates.

The RNN block is a component that searches for optimal language type information and m-th segment text information by conducting multiple simulations for various states and various actions based on a plurality of agents calculated in the policy network and value network. The RNN block according to an embodiment of the present invention, for example, can utilize Monte Carlo tree search, each node of the tree has a state, and each edge is expected according to a specific action for that state. It is a structure in which a leaf node is expanded whenever a transition is made to a new state by taking a new action with the current state as the root node. Searching for optimal language type information and m-th segment text information in the RNN block according to an embodiment of the present invention can be processed in four steps: Selection, Expansion, Evaluation, and Backup when Monte Carlo tree search is used.

The Selection step of the RNN block is a step in which the action with the highest value among selectable actions is selected and proceeded until a leaf node emerges from the current state. At this time, the value of the value function stored in the edge and the visit frequency value to balance exploration and use are used. The equation for selecting an action in the selection step is as follows.

In Equation 5 above, a _t is the action at time t (selection of language type information and m-th segment text information), Q (s _t , a) is the value of the value function stored in the tree, and u (s _t , a) is a value that is inversely proportional to the number of visits of the state-action pair, and is used to balance exploration and use.

The Expansion step of the RNN block is a step of adding a new node as a leaf node by acting according to the probability of the policy network learned through supervised learning when the simulation progresses to the leaf node.

The evaluation stage of the RNN block proceeds until the end of the value (reward potential) determined using the value network from the newly added leaf node and the language type information and the m-th segment text information selection episode using the policy network from the leaf node. This step evaluates the value of the leaf node through The equation below is an example of evaluating the value of a new leaf node.

In Equation 6 above, V(s _L ) is the value of the leaf node, λ is the mixing parameter, v _θ (s _L ) is the value obtained through the value network, and z _L may mean the compensation obtained by continuing the simulation.

The backup step of the RNN block is a step of reevaluating the value of the nodes visited during the simulation by reflecting the value of the newly added leaf node and updating the visit frequency. The equation below is an example of node value revaluation and visit frequency update.

In Equation 7 above, s _L ⁱ represents the leaf node in the ith simulation, 1(s,a,i) indicates whether the connection (s,a) was visited in the ith simulation, and when the tree search is completed, the algorithm It can be configured to select the most visited connection (s,a) from the root node. According to the RNN block according to an embodiment of the present invention, a plurality of language type information and m-th segment text information selected by a policy network are simulated multiple times based on a value network, and optimal language type information and m-th segment text information are preceded. An effect of being able to select segmental text information is generated.

According to one embodiment of the present invention, the reinforcement learning module may be configured such that a plurality of agents are configured. When a plurality of agents are configured, the language type information and the m-th segment text information selected by the reinforcement learning module compete with each other for a specific state, specific language type information, and m-th segment text information, respectively, to obtain the most optimal language type information and m-th segment text information. An effect of being able to select segmental text information is generated.

13 is a flowchart illustrating an operation example of a reinforcement learning module according to an embodiment of the present invention.

As shown in FIG. 13, when state s(t) is input by the text sequence generating neural network module, various language type information and m-th segment text information are generated by a plurality of agents of the policy network through the value network. The language type information and the m-th segment text information are selected by the probability a(t) of the selected language type information and the m-th segment text information, which are input to the RNN block and are actions output by the RNN block. t ends and episode t+1 begins. In episode t+1, s(t+1), a state change by a(t), is input by the text sequence generating neural network module, and r(t+1), a reward according to a(t), is immediately input, Value and policy networks will be updated.

The text information generating module may refer to a module that uses the text coordinate information and the text sequence information as input data and text information configured in the data format of [coordinates, text] as output data. For example, text information for a specific text box may have a structure of [(x,y), 'text'].

14 is a schematic diagram illustrating a text processing module according to an embodiment of the present invention. As shown in FIG. 14, the text processing module of the artificial intelligence-based electronic document management device 1 inputs full text information, which is information in which the text information included in the electronic document image is sequentially integrated in the order of y coordinates It is a module that receives and outputs item information that matches text information included in an electronic document for each item. A text processing module according to an embodiment of the present invention may be configured to include an electronic document classification neural network module, an item classification neural network module, and an item information generation module. For example, full text information may be configured in the form of [(20,2) real estate sales contract, (2,4)1. display of real estate, (2,8) location...].

Regarding the electronic document classification neural network module, FIG. 15 is a schematic diagram illustrating an electronic document classification neural network module according to an embodiment of the present invention.

As shown in FIG. 15, the electronic document classification neural network module takes as input data all text information, which is information in which the text information included in the electronic document image is sequentially integrated in the order of y coordinates, and the electronic document type class and each It can be configured as an artificial neural network module that outputs electronic document type information (eg, [real estate sales contract, 0.85]) including confidence in class. Regarding the learning session of the electronic document classification neural network module, the training session of the electronic document classification neural network module is such that the parameters of the electronic document classification neural network module are updated by the client learning module based on the difference between the output electronic document type information and the ground truth. can be configured.

Regarding the item classification neural network module, FIG. 16 is a schematic diagram illustrating an item classification neural network module according to an embodiment of the present invention.

As shown in FIG. 16, the item classification neural network module takes the text to be classified, which is the text subject to item classification among the text information, and the full text information as input data, and includes the classified item class and confidence. It can be composed of an artificial neural network module that outputs item type information (for example, [property location, 0.85]) as output data. Regarding the learning session of the item classification neural network module, the learning session of the item classification neural network module may be configured so that parameters of the item classification neural network module are updated by the client learning module based on the difference between the output item type information and the ground truth. . At this time, the item classification neural network module infers all possible item classes regardless of the electronic document type class, and the output item type information can be configured to be multiplied by the item vector for each electronic document, and the item vector for each electronic document is It means a preset vector to which a null value ([0]) is applied to items not included according to the electronic document type class of type information. For example, since the item classification neural network module infers all possible item classes regardless of the electronic document type, even if the electronic document type class is [Real Estate Sales Contract], for a specific classification target text, for the [Cheonse Deposit] class It can be configured to infer a confidence score. At this time, the item vector for each electronic document preset in the electronic document type class of [Real Estate Sales Contract] is matched with [0] for [Cheonse Deposit], and the item vector for each electronic document is applied to the inferred item type information [Cheonse The final confidence for the deposit] may be configured to be inferred as [0]. According to this, a separate item classification neural network module is not configured according to the electronic document type class, and item classification for a plurality of electronic document type classes is possible with only one item classification neural network module.

The item information generation module receives the classification target text and the item class (item type information) from the item classification neural network module, and generates item information for each item class by integrating other classification target texts classified into the same item class. am. For example, if the text to be classified is [Gangnam-gu, Seoul ...] and the item class of item type information is [Location of a property], the item information is a data structure in the form of [Location of a property, Gangnam-gu, Seoul ...] can be configured. In addition, for example, the text to be classified that is classified as the item class of [Location of property] is [Gangnam-gu, Seoul ...] (text to be classified) and [Samseong-dong Hillstate 2 Danji 2**-2*** Building] ( If there are two or more different texts to be classified), item information may be generated by integrating the text to be classified with other texts to be classified.

Regarding the neural network processing module, FIG. 17 is a schematic diagram showing a neural network processing module according to an embodiment of the present invention.

As shown in FIG. 17, the neural network processing module updates the parameters by processing the text extraction module and the learning session of the artificial neural network module included in the text processing module, and uploads the updated parameters to the federated learning module of the electronic document management server. It is a module that downloads the main neural network jointly learned by a plurality of clients in the main neural network module of the electronic document management server. A neural network processing module according to an embodiment of the present invention may include a correction information generation module, a client learning module, a parameter upload module, and a main neural network download module.

Regarding the correction information generation module, FIG. 18 is a schematic diagram showing the correction information generation module according to an embodiment of the present invention.

As shown in FIG. 18, the correction information generation module outputs the electronic document image, the electronic document type information, the text coordinate information, and the item information to a display of a user client, and outputs information about the information by a user's input. This is a module that receives the correction information reflecting the user's correction. For example, if the electronic document type information is inferred as [real estate sales contract] and output on the display of the user client, the user checks the electronic document image and modifies the electronic document type information to [real estate lease contract] to display the corrected information. can be configured to create In this case, when the electronic document type information is modified, since the item vector for each electronic document is changed, item type information and item information may be changed. In addition, for example, if the text coordinate information is inferred as [12,131,5,1,1] and output on the display of the user client, the user checks the electronic document image and converts the text coordinate information to [12,132,6,2,2] It may be configured to generate correction information by modifying to . In addition, for example, when item information is inferred as [Location of sale, Gangnam-gu, Seoul...] and output on the display of the user client, the user checks the electronic document image and returns the item information to [buyer's address, Gangnam-gu, Seoul...] It may be configured to generate correction information by modifying to .

In addition, the correction information generating module may be configured to further include an abnormal text classification module that infers whether or not the text subject to classification of the item information is abnormal. When the text to be classified is inferred to be abnormal, it may be displayed on the display of the user client so that the user can easily generate correction information.

The abnormal text classification module is a module that receives a multi-dimensional vector incorporating item type information and classification target text of item information and classifies whether or not the classification target text of item information is abnormal text. Specifically, the abnormal text classification module according to an embodiment of the present invention may include a generation module and a classification module, and the generation module receives random noise (Z) using the classification module and receives normal item type information and classification target. After learning the generation module to generate a normal vector, which is a multidimensional vector that integrates text, the loss function output value L (Loss) of the generation module is used as an abnormal text score according to the change in random noise (Z) input to the generation module. Whether or not data input to the abnormal text classification module is abnormal text can be distinguished based on whether L is lowered to a specific value or less.

The generation module of the abnormal text classification module according to an embodiment of the present invention may be configured to generate a normal vector composed of an encoder and a decoder, and the encoder of the generation module is a multidimensional vector in which normal item type information and text to be classified are integrated. It can be composed of a plurality of consecutive ConvNets that receive and encode into latent variables of 1 x 1 x k, and the decoder of the generating module transforms latent variables of 1 x 1 x k into multidimensional vectors including item type information and text to be classified. It can be composed of a plurality of consecutive networks that decode to output. In this case, the generation module may be trained to input a multidimensional vector that is a normal vector and output a multidimensional vector close to the normal vector, and may be learned by a discrimination module that distinguishes whether or not the multidimensional vector output by the generation module is a normal vector. there is. The classification module of the abnormal text classification module according to an embodiment of the present invention may be configured to distinguish whether a multidimensional vector output by the generation module is a normal vector through a CONCAT function and a plurality of encoders.

Regarding learning of the abnormal text classification module, FIG. 18 is a schematic diagram showing a learning process of the abnormal text classification module according to an embodiment of the present invention. As shown in FIG. 18, in the generation module, a Loss Function may be configured to configure a classification module and a MinMax game, and may be learned at the same time. Equation 8 below is a loss function of the generation module and the classification module.

In Equation 8 above, G is a generation module, D is a classification module, z is random noise input as a latent variable, y is a normal vector that is a multidimensional vector integrating normal item type information and text to be classified, G (x ) means a generated vector, which is a generated multidimensional vector. Therefore, according to Equation 8, the loss function of the generation module and the classification module is D if the generation module is not sufficiently learned and the classification module perfectly distinguishes y and G (z) through random noise z, which is a latent variable. (z, y) = 1, with a max value of 0 by D (z, G (z)) = 0, and after learning the generation module, the discrimination module distinguishes y and G (z) through random noise z If not, it has a min value of -log4 by D(z,y)=1/2 and D(z,G(z))=1/2. That is, when the generation vector G(z), which is a multidimensional vector generated by the generation module through random noise z by the above loss function, and the normal vector y, which is a multidimensional vector integrating normal item type information and text to be classified, are the same, The generation module has a global minimum, and in this direction, the generation module and the classification module are learned. The effect of quickly optimizing the generation module is generated by the mutual antagonistic dependency between the generation module and the division module.

Regarding the abnormal text classification of the abnormal text classification module, FIG. 19 is a schematic diagram showing an abnormal text classification process of the abnormal text classification module according to an embodiment of the present invention. As shown in FIG. 19, the generating module receives random noise z to generate a multi-dimensional vector (item type information and text to be classified as a generated vector) close to a normal vector, and the input item type information and text to be classified are Based on the difference between the multi-dimensional vector y (a classification target vector that determines whether or not it is abnormal text) and the generated vector, whether the input item type information and the text to be classified is normal or abnormal is distinguished.

In addition, the generation module according to an embodiment of the present invention may be configured so that the parameter is fixed after learning, and the latent variable is It can be configured to adjust the random noise z. Equation 9 below relates to a division loss function (L) for the difference between G(z) and y, and Equation 10 relates to control of random noise, which is a latent variable.

In Equations 9 and 10 above, L is the division loss function that is the difference between the generation vector generated close to the normal vector and the division target vector (input item type information and classification target text), G (z) is the generation vector, and z is the latent Random noise as a variable, y is a vector to be classified, and η is a learning rate. According to this, if the multidimensional vector y (classification target vector) of the item type information and the text to be classified is normal text, if z is adjusted to reduce the division loss function L with the parameters of the generation module G fixed, the loss value of L is lowered below a certain value. In addition, if the multidimensional vector y (vector to be classified) of the user data (data to be classified) is abnormal text, even if z is adjusted to reduce the division loss function L with the parameters of the generation module G fixed, the loss value of L is specific. It does not fall below the value. That is, Loss L when y is abnormal text has a relatively higher value than L when y is normal text. Therefore, an effect of being able to classify abnormal text, classify abnormal text, and detect abnormal text using L as an anomaly score is generated.

In addition, the correction information generation module, if it is inferred as abnormal in the abnormal text classification module that infers whether or not the text subject to classification of item information is abnormal, generates normal text for the abnormal text and presents it to the user client. It may be configured to include more. When the text to be classified is inferred to be abnormal, the normal text generated by the normal text generation module may be displayed on the display of the user client so that the user can easily generate correction information.

Regarding the configuration of the normal text generation module, FIG. 20 is a schematic diagram showing the structure of the normal text generation module according to an embodiment of the present invention. As shown in FIG. 20, Downsampling & Upsampling in the form of a combination of an encoder and a decoder for outputting normal text as output data, using a combination vector obtained by concatenating item type information and text to be classified as input data of the normal text generation module. It can be composed of an artificial neural network.

Regarding the overall embodiment of the normal text generation module, for example, an encoder according to an embodiment of the present invention receives item type information and text to be classified as input data and generates an encoding vector, which is a latent variable of 1 x 1 x k It can be composed of a ConvNet including a plurality of consecutive Convolution Layers, Pooling Layers, and Fully Connected Layers that encode output data. In addition, the encoder may be composed of a decoder and a skip connection structure. In the learning session of the encoder, item type information and text to be classified input to the encoder according to an embodiment of the present invention may be configured to be input in a channel-wise concatenation structure for each convolution layer of the encoder and decoder. At this time, the vanishing gradient is improved and feature propagation is strengthened in the learning session of the encoder and decoder by the configuration in which the item type information and the text to be classified are input in the structure of channel-wise concatenation for each convolution layer of the encoder and decoder, The number of parameters is saved, resulting in a reduction in computing resources. In the learning session of the normal text generation module, parameters of the normal text generation module may be updated in a direction to reduce a loss consisting of a difference between the normal text and the corresponding reference data (Ground Truth). The loss function of this normal text generation module can be composed of mean square loss, cross entropy loss, and the like.

The client learning module 40, when the correction information generation module receives correction information reflecting the user's correction, uses the correction information as the ground truth to update the parameters of the artificial neural network included in the text extraction module and the text processing module. You can conduct learning sessions of text extraction module and text processing module.

Regarding the learning session of the text box generating neural network module, which is a component of the text extraction module of the client learning module 40, the learning session of the text box generating neural network module of the client learning module 40 is received from the correction information generating module. It can be configured to be performed when the correction information includes correction for the text box information, and the text box, which is a bounding box that inputs the electronic document image as input data to the text box generating neural network module and surrounds the text included in the electronic document image The text box information including the coordinates, height, width, and angle of is used as output data, and the loss of the output text box information and the modified text box information (ground truth) of the correction information is reduced (or , in a direction in which similarity increases) parameters of the text box generating neural network module may be updated.

Regarding the learning session of the electronic document classification neural network module, which is a component of the text processing module of the client learning module 40, the learning session of the electronic document classification neural network module of the client learning module 40 is received from the correction information generating module. It can be configured to be performed when the correction information includes modification of the electronic document type class, the electronic document classification neural network module takes text information as input data, and the electronic document type class and weight for each item as output data. , The parameters of the electronic document classification neural network module are updated in the direction in which the loss of the output electronic document type class and the modified electronic document type class (ground truth) of the correction information becomes smaller (or in the direction of similarity increases). It can be configured so that

Regarding the training session of the item classification neural network module, which is a component of the text processing module of the client learning module 40, the learning session of the item classification neural network module of the client learning module 40 is the correction information received from the correction information generation module. It can be configured to be performed when modification of the item class is included in the item classification neural network module, and a pair of the text to be classified, which is the target text for item classification among the text information, and the full text information are input data, and the classification is performed. The output data is the item class and confidence, and the loss of the item class output for a specific classification target text and the modified item class (ground truth) of the correction information is reduced ((or, In a direction in which similarity increases) parameters of the item classification neural network module may be updated.

After the learning session of the text extraction module and the text processing module by the client learning module 40, the parameter upload module 50 transfers the changed parameters of the text extraction module and the text processing module to the combined learning module of the electronic document management server 200. This module is uploaded to (220). The parameters may include the gradient (g) or the weight (w) of the artificial neural network after the learning sessions of the text extraction module and the text processing module. Parameter upload to the text box generating neural network module of the parameter upload module 50 is configured to be performed when the correction information received from the correction information generation module includes correction of the text box information, and the parameter upload module 50 Parameter upload for the electronic document classification neural network module is configured to be performed when the correction information received from the correction information generation module includes correction for the electronic document classification class, and the item classification neural network module of the parameter upload module 50 Parameter upload for is configured to be performed when modification information received from the modification information generating module includes modification of an item class.

In addition, the parameter upload module 50 may be configured to apply noise ε to the changed parameters of the text extraction module and the text processing module and upload them to the electronic document management server 200 . For example, if the parameter is the gradient (g), it can be configured to be uploaded as g+ε, and if the parameter is the weight (w), it can be configured to be uploaded as w+ε. In this case, the noise ε may be configured to have a positive value and a negative value randomly assigned so that the effect of the noise is minimally applied during federated learning in the federated learning module 220 . According to this, since noise is applied to the parameter in the client and uploaded to the electronic document management server, even if a third party acquires the parameter, it is impossible to obtain the electronic document information.

The main neural network download module 60 downloads the text box generation main neural network, the electronic document classification main neural network, and the item classification main neural network pretrained by the federated learning module 220 of the electronic document management server 200 to generate a text box. This is a module that replaces (replaces, transfers) at least some networks of the neural network module, the electronic document classification neural network module, and the item classification neural network module.

The network download of the text box generating neural network module downloads the pretrained text box generating main neural network from the main neural network module 210 of the electronic document management server 200, and the back layer of the text box generating neural network module (professional style discriminator) It is configured to replace (replace, transition) the rest except for the last n layers learned by (130) with the downloaded text box generating main neural network. According to this, since the neural network of each client's text box generating neural network module is not completely replaced with the main neural network, it is possible to maintain the style of a professional text box output from each client while continuously updating the text box generating neural network module. The effect of enabling personalization of the text box is generated.

Network download of the electronic document classification main neural network downloads the electronic document classification main neural network pretrained in the main neural network module 210 of the electronic document management server 200, and downloads the electronic document classification neural network module. It is configured to be replaced (replaced, transferred) with.

Network download of the item classification main neural network downloads the pre-learned item classification main neural network from the main neural network module 210 of the electronic document management server 200, and replaces the item classification neural network module with the downloaded item classification main neural network. , transition).

The electronic document management server 200 may include a main neural network module 210 and a federated learning module 220, a text box generation neural network module uploaded from a plurality of user clients 100, an electronic document classification neural network module, A server configured to collect parameters of the item classification neural network module, update the main neural network module 210, and then redistribute the pretrained main neural network to the user client 100.

The main neural network module 210 may include a text box generating main neural network, an electronic document classification main neural network, and an item classification main neural network. Parameters can be configured to be updated. The text box generating main neural network refers to the main neural network corresponding to the text box generating neural network module, and the electronic document classification main neural network refers to the main neural network corresponding to the electronic document classification neural network module, and the item classification main neural network refers to the item classification. It means the main neural network corresponding to the neural network module.

The federated learning module 220 collects the parameters of the text box generating main neural network, the electronic document classification main neural network, and the item classification main neural network uploaded from the plurality of user clients 100, and then uses the collected parameters as the main neural network module. (210) is a module configured to update. At this time, the method of updating the main neural network module 210 using the collected parameters is the difference between the correction information received from the correction information generation module and the information before correction (text box information, electronic document type class, item class) as follows The correction change information may be used as a weight of the parameter, and the parameter may be averaged based on the correction change information by dividing the sum of the product of the parameter and the correction change information by the sum of the correction change information.

When the parameter is the gradient (g), the federated learning module 220 may be configured to update the parameter of the main neural network module 210 as shown in the following equation.

In Equation 11 above, w _new is the weight after updating the main neural network module 210, w _old is the weight before updating the main neural network module 210, α is the learning rate, and s _n is the user client 100 Upload from n The modified change information, g _n , may be configured to mean a gradient uploaded from the user client 100 n .

When the parameter is weight (w), the federated learning module 220 may be configured to update the parameter of the main neural network module 210 as shown in the following equation.

In Equation 12 above, w _new is the weight after updating the main neural network module 210, s _n is the correction change information uploaded from the user client 100 n, and w _n is the weight uploaded from the user client 100 n. can be configured to mean

According to this, during joint learning in the main neural network module 210, correction change information is used as a weight of a parameter, thereby generating a mini-batch effect. In addition, during federated learning in the main neural network module 210, correction change information is used as a weight of a parameter, thereby reducing network topology and asynchronous communication problems. In addition, when the difference between the corrected information and the pre-corrected information is large, an effect of having a greater influence on the update of the main neural network module occurs.

21 schematically illustrates an artificial intelligence-based electronic document management device having an automatic notification function according to an embodiment of the present invention.

Referring to FIG. 21, the artificial intelligence-based electronic document management device 1 having an automatic notification function receives an electronic document image and outputs a plurality of items of information obtained by matching text included in the electronic document image for each item. can To this end, a text extraction module including an artificial neural network that receives an electronic document image, extracts text included in the electronic document image, and outputs text information, and the entire text, which is information in which the text information is sequentially integrated in the order of y coordinates A text processing module including an artificial neural network that receives information and outputs item information obtained by item-by-item matching of texts included in electronic document images, and processes learning sessions of the artificial neural network included in the text extraction module and text processing module. One or more artificial neural network modules including a neural network processing module for updating parameters and an automatic notification module for determining automatic notification target item information to be automatically notified among a plurality of item information may be used.

Among the one or more artificial neural network modules, the text extraction module, the text processing module, and the neural network processing module have been described in detail above, and therefore, redundant descriptions will be omitted.

The automatic notification module according to an embodiment of the present invention may select and determine automatic notification target item information requiring automatic notification among a plurality of items of information. For example, when the target document of the electronic document image is a real estate sales contract, the plurality of item information generated by the artificial intelligence-based electronic document management device 1 is, for example, the location of the property, the contract amount, and the parties to the contract , the date of the contract, the date of balance, and the renewal date of rent on a monthly basis. Among them, balance date and monthly rent renewal date may correspond to item information that can be helpful to the user if automatic notification is performed, so the automatic notification module provides automatic notification target items through reinforcement learning-based artificial neural network module. Information can be selected and determined. And, the automatic notification module 2 may perform automatic notification to a user device associated with an electronic document, or more preferably, associated with automatic notification target item information. Such automatic notification may be transmitted using a notification function of the user client or in various ways such as SNS, messenger message, and e-mail, but is not limited thereto.

22 schematically illustrates an automatic notification module according to an embodiment of the present invention.

Referring to FIG. 22, the automatic notification module 2 includes a selection module for selecting item information related to a date from among a plurality of item information generated by the artificial intelligence-based electronic document management device 1, and related to the selected date. An automatic notification item determination module for determining item information to be automatically notified among item information may be included.

The selection module may determine whether each of the plurality of items of information is related to a date. For example, in the case of a date, since it can be written in various notation methods such as October 9, 2022, 2022.10.9, 22/10/9, etc., the screening module determines whether each of the plurality of item information includes a number and includes a number If so, you can determine whether a number represents a date. And, the selection module may transmit the item information related to the date (ie, item information related to the date) to the automatic notification item determination module.

The automatic notification item determination module may determine item information to be automatically notified from among one or more date-related item information. In this case, the automatic notification item determination module may determine automatic notification execution target item information on a rule-based basis. For example, item information such as "balance date", "intermediate payment payment date", and "expiration date of rent for cheonsei" input in advance by the manager may be determined as item information subject to automatic notification.

According to another embodiment, the automatic notification item determination module may determine automatic notification execution target item information from among one or more date-related item information using a reinforcement learning-based artificial neural network module.

23 is a diagram schematically illustrating a transformer-based artificial neural network module included in an automatic notification item determination module according to an embodiment of the present invention.

Referring to FIG. 23, the automatic notification item determination module may include a transformer-based artificial neural network module, and through reinforcement learning using the transformer-based artificial neural network module 10, automatic notification target item information (a _t ) including Optimum output can be produced.

The transformer-based artificial neural network module 10 includes a future expected reward (R _t ), state information received by the automatic notification item determination module (S _t ), and an automatic notification target item that is a result previously output from the artificial neural network module 10 Past output (a _t-1 ) for information can be input.

Here, the future expected reward (R _t ) is expected in the future by taking a specific action in the automatic notification item determination module as Return to go, that is, by the automatic notification item determination module generating automatic notification target item information (a _t ). represents the sum of the rewards. For example, the automatic notification item determination module may be designed to give different rewards depending on whether an item actually requires automatic notification based on a user's feedback when the automatic notification is performed.

Status information (S _t ) is information received by the automatic notification item determination module, and may include date-related item information and, optionally, related law information. As shown in FIG. 22, the legal information may be separately input into the automatic notification item determination module in the latest version (ie, by reflecting the amendment).

The past output (a _t-1 ) is an automatic notification target item that is a result previously output from the transformer-based artificial neural network module 10 when input and output to the transformer-based artificial neural network module 10 is performed in real time or time series. Indicates the past output (a _t-1 ) for information.

In this way, when future expected compensation (R _t ), state information (S _t ), and past output management mode (a _t-1 ) are input to the transformer-based artificial neural network module 10, the transformer-based artificial neural network module 10 automatically In the notification item determination module, automatic notification target item information (a _t ), which is the optimal output to be selected at the present time, can be generated and output.

Hereinafter, a process of processing input data through the transformer-based artificial neural network module 10 and outputting a resultant value will be described.

Referring to FIG. 23, according to an embodiment of the present invention, the transformer-based artificial neural network module 10 for generating automatic notification target item information (a _t ) of the automatic notification item determination module includes a first block 10a (eg For example, an encoder) and a second block 10b (eg, a decoder).

An expected future reward (R _t ) and state information (S _t ) may be input to the first block 10a of the artificial neural network module 10 . In addition, the future expected compensation (R _t ) and state information (S _t ) input in this way are input to the embedding module of the first block 10a and vectorized so that embedding information is output and a bundle of output embedding information is divided into different linear layers Through linear embedding, it is composed of Q (Query feature), K (Key feature), and V (Value feature) including features dimensions, and then Q (Query feature), K (Key feature), and V (Value feature) are multi - Can be used as the input data of the head Attention Layer. Q, K, and V may be generated through different linear embeddings from embedding information. Q and K are input as input data of the 1st MatMul operation, calculate the similarity for all K for Q through Scale and softmax operation, and output a similarity vector between a series of embedding information, and this similarity vector and V are 2nd MatMul It can be input as input data of operation. In addition, the multi-head attention layer can output attention information based on the input Q, K, and V values and input it to the feed forward layer. These multi-head attention layers and feed forward layers can be repeated N times, and the encoding information output from the previous feed forward layer can be repeated in such a way that the input of the next multi-head attention layer. And the last Feed Forward Layer can output final encoding information, which is K (Key), V (Value) It can be input as a value.

In addition, the past output (a _t−1 ) may be input to the second block 10b of the transformer-based artificial neural network module 10 . Here, the past output (a _t-1 ) is input to the embedding module of the second block 10b and vectorized to output embedding information, and linear embedding a bundle of output embedding information through different linear layers to include features dimension. After configuring Q (Query feature), K (Key feature), and V (Value feature), Q (Query feature), K (Key feature), and V (Value feature) are used as input data of the first multi-head attention layer. can do. And the first multi-head attention layer can output a Q (Query) value, and this Q value is input to the second multi-head attention layer. The second multi-head attention layer may receive final encoding information from the first block 10a as K and V values. And the second multi-head attention layer can output attention information based on the input Q, K, and V values and input it to the Feed Forward Layer. The first multi-head attention layer, the second multi-head attention layer, and the feed forward layer may be repeated N times, and the decoding information output from the previous feed forward layer is input to the next first multi-head attention layer. Iterations can be made in this way. And the last Feed Forward Layer can output final decoding information, and this final decoding information goes through a Linear Layer and a Softmax Layer, and the Transformer Artificial Neural Network module 10 is the optimal output to be selected at this time in the automatic notification item determination module. Automatic notification target item information (a _t ) can be created.

However, a specific implementation method of the transformer-based artificial neural network module 10 as described above may be variously changed, but is not limited thereto.

As described above, those skilled in the art to which the present invention pertains will be able to understand that the present invention can be embodied in other specific forms without changing its technical spirit or essential features. Therefore, the above-described embodiments should be understood as illustrative in all respects and not restrictive. The scope of the present invention is indicated by the claims to be described later rather than the detailed description, and all changes or modifications derived from the meaning and scope of the claims and equivalent concepts should be construed as being included in the scope of the present invention.

The features and advantages described in this specification are not all inclusive, and many additional features and advantages will become apparent to those skilled in the art, particularly from consideration of the drawings, specification, and claims. Moreover, it should be noted that the language used herein has been chosen primarily for readability and instructional purposes, and may not have been chosen to delineate or limit the subject matter of the invention.

The foregoing description of embodiments of the present invention has been presented for purposes of illustration. It is not intended to limit the invention to the precise form disclosed or to do so without omission. Those skilled in the art can appreciate that many modifications and variations are possible in light of the above disclosure.

Therefore, the scope of the present invention is not limited by the detailed description, but by any claims of the application based thereon. Accordingly, the disclosure of embodiments of the invention is illustrative and not limiting of the scope of the invention set forth in the claims below.

Claims (5)

An artificial intelligence-based electronic document management device with an automatic notification function,
one or more artificial neural network modules for receiving an electronic document image and outputting a plurality of items of information obtained by item-by-item matching of text included in the electronic document image;
a selection module for selecting item information related to a date from among the plurality of item information; and
Automatic notification item determination module for determining automatic notification target item information to be automatically notified among the selected item information
Including,
The automatic notification item determination module,
A transformer-based artificial neural network module that receives the selected item information and outputs item information for automatic notification,
The transformer-based artificial neural network module receives future expected compensation, state information, and past outputs of the artificial neural network module, and generates an optimal output including automatic notification target item information as an output,
The future expected reward represents the total amount of rewards expected in the future as the automatic notification item determination module generates the automatic notification target item information;
The status information includes at least the selected item information received by the automatic notification item determination module;
The past output includes automatic notification target item information output in the past by the transformer-based artificial neural network module.
Artificial intelligence-based electronic document management device with automatic notification function.
delete According to claim 1,
The status information includes legal information related to the electronic document image.
Artificial intelligence-based electronic document management device with automatic notification function.
According to claim 1,
The transformer-based artificial neural network module includes a first block and a second block,
The first block receives the future expected compensation and the state information, outputs final encoding information, and provides it to the second block as an input;
The second block receives the past output and generates an optimal output including the automatic notification target item information.
Artificial intelligence-based electronic document management device with automatic notification function.
According to claim 1,
The one or more artificial neural network modules,
a text extraction module including an artificial neural network that receives an electronic document image, extracts text included in the electronic document image, and outputs text information; and
A text processing module including an artificial neural network that receives full text information, which is information in which the text information is sequentially integrated in the order of y coordinates, and outputs item information obtained by item-by-item matching of the text included in the electronic document image.
An artificial intelligence-based electronic document management device that includes an automatic notification function.