Structuring natural text entities using neural networks for generating 3D-scenes

The subject of this study is the automation using neural networks of the process of assembling three-dimensional scene, which can be used both to generate three-dimensional scenes or locations in computer games from a textual description, and to prepare sequences of three-dimensional synthetic data. This topic is relevant for developing three-dimensional graphics including interactive projects – games, simulators, AR/VR applications. After analyzing and comparing the results obtained in a number of well-known completed projects, the authors determine technologies and software libraries, which allow effectively achieving the desired goal - to provide fast assembly of three-dimensional scenes filled with objects according to the text description. Thanks to synthesis of the best solutions, it was possible to create an optimal concept that allows achieving quick and qualitative result with the right rules of building geometrical relations between scene objects. There is a formed list of requirements to the designed tool and its architecture. Input data for using this tool is a text in natural language; output data is a scene with objects corresponding to the description used. The main result achieved is a finished software tool for Unreal Engine developed on the basis of the nlpue4 neural network and the set of tensorflow, nltk, pandas, gensim, h5py libraries. The readiness of the tool is evaluated as a prototype solution, which can be integrated into the drafting stage of interactive projects with three-dimensional graphics. To evaluate the created tool effectiveness objectively, the authors have conducted the experiments that proved that its use even in the current version significantly reduces development time and does not require a user to have skills in programming or creating three-dimensional graphics. There is also a discussion about the research development prospects.