METHOD FOR AUTOMATIC EXTRACTION OF TEXT INFORMATION FROM IMAGE FILES

Description

[0001] The present invention relates to a method for automatically extracting information from document files according to the preamble of claim 1. It falls within the field of digital document processing and optical text recognition.

[0002] Document data extraction systems have undergone significant improvements in the last 10 years. They are used to read any type of document, whether paper documents, electronically generated documents, or purely electronic formats, and to transform the extracted data into a structured format usable by other business systems.

[0003] The aim is to create structured data from unstructured or semi-unstructured data. This is generally attempted using two different approaches: template-based approaches and freeform recognition. Template approach

[0004] The template approach creates a template for each document layout. This template contains the coordinates of the values ​​to be searched for or predefined anchor points, thus automatically inserting the values ​​into the document structure when it is read. While this method works well in principle, it is complex to set up and therefore limits its application. Furthermore, the template approach cannot handle layouts that are new to the system. Freeform recognition

[0005] Freeform recognition attempts to structure document values ​​using various extraction rules, without templates. This method works well with a limited set of document variants, as a rule can be defined for each case. However, setting up the rules is complex and prone to errors. The freeform recognition approach handles new documents better than the template approach, but even here, significant additional configuration is usually required for new document layouts. Such systems are often overwhelmed, especially with documents containing previously unknown information.

[0006] Machine learning algorithms are increasingly used in document extraction with both approaches, and can significantly improve extraction results, especially when rules and templates are cleverly used and, if necessary, combined. These machine learning algorithms are each tied to / specialized for specific document types. This means that at least one machine learning algorithm must be available for each document type. Fig 1 ).

[0007] One possible approach is client-specific learning. Here, the learning models are usually trained universally, but in actual use, they are limited to the client. This means that further learning only takes place within the document data processed by that client. If the same system is used by different clients, the same, slightly modified learning occurs in parallel. This learning increases the cost of system setup and significantly extends the time-to-solution for the client.

[0008] Pure machine learning approaches have also become more widespread in recent years. These attempt to solve document extraction using a purely data-driven methodology. These systems are delivered with basic learning for each document type and are subject to the same limitations of client-specific learning as learning systems using template and free-form approaches.

[0009] US 9,384,264 B1 discloses a machine-learning-based approach for classifying documents and extracting information from documents.

[0010] The present invention aims to introduce a new method that utilizes known methods by attributing learning models to extraction entities (instead of document types). This allows all users to benefit from continuous improvements in extraction methods and models, thus enabling the cost-effective processing of very small quantities.

[0011] This problem is solved by a method with the features of claim 1. Further features and embodiments are evident from the dependent claims and their advantages are explained in the following description. Drawings

[0012] Fig. 1 Document extraction of conventional systems (state of the art) Fig. 2 Learning per extraction entity across document types Fig. 3 Document Swarm Learning: "Learning Flow"

[0013] The figures represent possible embodiments, which are explained in the following description. Description

[0014] The inventive process comprises the following sub-processes: Attribution of learning models to extraction entities

[0015] Instead of binding the learning models to document types as in conventional systems, the new method divides the models into extraction entities ( Figure 2 Extraction entities represent a structured, normalized value to be retrieved. In a software application, this could be, for example, a data field. Fig 2 Three possible entities are shown, which are arranged in different positions in the various document types.

[0016] This method is used independently of the actual extraction approach (see sections 2.1 to 2.4).

[0017] In this process, a) the extraction entities are separated from the document types. b) the extraction entities are decoupled from the learning models and extraction methods and re-established relationships between them.

[0018] For each extraction entity, learning models and extraction mechanisms can be compared in competition with each other. In this case, a further model represents the evaluation of the best learning model and the best extraction method in each case.

[0019] In this way, it is possible to improve learning models independently of the document type. This improvement means that no additional training effort is required for the same extraction entities for each client. Furthermore, this method enables the simple and significantly faster configuration of new document types with "pre-trained" extraction entities. This sub-method also eliminates redundancies in the training of learning models compared to existing approaches. Fig 3 A second, complementary sub-method is shown, which applies the "attribution of learning models to extraction entities" to learning in a client network.

[0020] Each client in the system maintains a processing pipeline for the documents, which extends over the following 4 main steps: 1) Document input

[0021] Users upload documents to the platform via their preferred channel. Any channel is acceptable for uploading. 2) Classification the Document types & potential page split

[0022] The document is classified as a specific document type. Document separation is also performed as needed. Document separation becomes necessary when the user uploads multiple documents to the platform in a single file. Document classification and separation are carried out using the production classification models (see 11). 3) Document data extraction

[0023] In the next step, the document data is extracted. The extraction entities are configured in the document type. Data extraction is performed using the production extraction models (see 12). 4) Document output

[0024] Once the document data extraction is complete, the structured data is made available for consumption or export.

[0025] The following intermediate steps are used in the sequence of the 4 main steps: 5) and 6) Formal validation of the classification and extraction results

[0026] The process allows the user to review and correct the classification and extraction results. In doing so, the user collects the information necessary to improve the learning models. 7) Validation Data Gatekeeper

[0027] The Validation Data Gatekeeper is a component that checks the results of the formal validation and evaluates them according to objective criteria. The goal is to identify data that is not helpful for training as early as possible and exclude it from further processing. 8) and 9) Retraining machine learning models for classification and document data extraction

[0028] Using the data obtained through the functional validation and checked by the Validation Data Gatekeeper, the models of the extraction entities are granularly reconstructed according to the learning of the extraction entities. Fig 2 trained. 10) Model Quality Gatekeeper

[0029] The Model Quality Gatekeeper is a component that compares the performance of trained (historical and competing) models. If one or more models per extraction entity are rated as performing better, they are replaced in the current, production environment (see Figures 11 and 12). 11) and 12) Productive extraction models

[0030] The production models are continuously used for extraction in the document processing pipeline. They form the basis for reading document data for live operation.

[0031] A "swarm effect" is achieved by sharing the improved models with all users on the platform. Similarly, all data from the formal validation following quality assurance is incorporated into the improvement of the learning models. All improvements to the learning models are then made available to all users on the platform. The entire learning process is fully automated.

Claims

1. A method for automatically extracting information from a document file, wherein the document file contains at least one extraction entity in the form of a value to be found, a list of different recognizable document types is provided, at least one extraction entity is configured for each document type, each extraction entity is assigned a specific extraction model, and each extraction model is independent of the document type, comprising the following steps: • Recognition of the document type of the document file using a classification model; • Extraction of the information contained in the extraction entity in the document file using the specific extraction model assigned to the extraction entity configured for the recognized document type, wherein at least two different document types with at least one identical extraction entity are configured with the same assigned extraction model.

2. The method of claim 1, wherein an extraction model is retrained by machine learning based on a user's review and correction information.

3. The method of claim 2, wherein a validation data gatekeeper evaluates the relevance of the user's review and correction information for machine learning training and excludes irrelevant review and correction information.

4. The method of claim 2, wherein a model quality gatekeeper compares the performance of historical and competing trained extraction models for an extraction entity and assigns the most efficient extraction model to the extraction entity.

5. Method according to claim 2, wherein an extraction model is retrained based on review and correction information from a user in connection with an extraction entity of a document type, and the retrained extraction model is used in the context of extracting information from a document file of another user, wherein the document file is of a different document type but contains the same extraction entity.