The Federal Cadastral Chamber announced a tender for the creation of a modular machine learning system

2021: Tender for a Modular Machine Learning System

On September 23, 2021, it became known that the Federal Cadastral Chamber as part of the Federal Service for State Registration, Cadastre and Cartography (Rosreestr) is ready to spend up to 36 million rubles. to create a modular machine learning system (MSMO) as part of the optimization of state cadastral accounting and state registration of property rights.

This amount is set as the initial maximum contract price in the thematic tender of the Cadastral Chamber, which was launched on September 20, 2021 in the format of an open tender. Applications from applicants will be accepted until October 12, 2021, summing up is scheduled for October 18. The work front is designed for the period until December 15, 2021.

When developing the IMSS, software from the Unified Register of Russian Software under the Ministry of Digital Arts or related to open source programs should be used.

During the project, the contractor will have to create an intelligent system for supporting human decision-making using neural network technologies. Thus, the process of providing specialized public services of Rosreestr for citizens will be automated.

In particular, officials expect to minimize manual operations of users of the federal state information system of the unified state real estate registry (FGIS USRN) in the registration and registration area, reduce the number of possible errors associated with the performance of repeated routine user operations, reduce the processing time for documents when providing public services due to the use of neural network methods for processing natural language.

Also, according to the project, there should be an increase in the availability of information contained in incoming documents, due to the use of the computer vision mechanism, the transition to the format of machine-readable data at the stage of preliminary verification of documents, and the unification of the document processing process.

In accordance with the TA, the work on the creation of the MSMO should be carried out on a test copy of the USRN FGIS provided by the customer.

The incoming documents shall be classified in the IMS created (definition of their type). IMMS must perform normalization of image geometry - orient it correctly. The applied techniques shall operate with accuracy of determination of initial orientation exceeding 99.9%. Based on normalization of the image geometry of the scanned document, the segmentation algorithm should be improved. Recognition of different fonts shall be implemented in IMS.

In ISMO, neural network methods for processing the natural language should be applied, based on recent developments in the field of semantic analysis and allocation of named entities (NER), and austrimic methods (regular expressions) for compliance with accepted "masks," for example, series, passport number, cadastral number.

MISMO shall distinguish between the data of the entities of the registration action, real estate objects in respect of which registration actions are performed, as well as other material and other terms of the transaction in accordance with the developed algorithms from scan images of the submitted documents and application files in XML format.

The IMSS should implement a mechanism for obtaining information on readability checks, the absence of attributes and edits, the presence of signatures and/or seals, as well as a computer vision mechanism with the ability to visually determine entities in documents from a loaded package.

MISMO should provide automated comparison and analysis of compliance of document content with USRN information and information specified in the application received as part of the package of documents.

Based on the results of training on the minimum data set provided by the customer, the IMDS at the preliminary test stage should provide the value of the aggregate accuracy indicator - it is calculated as the ratio of the found units of named entities to the entire number of units of entities in the text - at the level of at least 60%.

MISMO should provide the possibility of improving the quality of data recognition through training during operation. According to the results of experimental operation, the contractor should provide the customer with a plan to gradually bring the total accuracy indicator of artificial intelligence neural networks to the target values ​ ​ of 70, 80, 90, 95% with the indication of the necessary volumes of formation of training data sets^[1].

Notes