Automatic location detection of the text

Efimova Valeria Alexandrovna

Over the past five years, deep learning models based on neural networks have achieved impressive results in the task of generating images from text. However, the images are generated with artifacts and still in insufficient resolution for printing. To correct this situation, we will divide the problem of image generation into subtasks, including determining the background of the image. From the text, you can try to understand what place is described or implied in it. This article proposes two methods for obtaining information about the place of action from a given text using natural language processing based on a pre-trained BERT transformer. The first method, called Location Extraction Transformer (LET), is designed to extract words from a text that explicitly mentions the place of action. The second method, called Location Inference Transformer (LIT), is designed to determine the location of an action that is implied in the text, but not directly mentioned. The performance of the proposed algorithms is compared by F1-measure with several existing approaches that can be used to extract information about the location of the text. Based on the results obtained during the comparison, it can be concluded that the proposed LET and LIT models turned out to be better than other algorithms.

Keywords:natural language processing; contextual image synthesis; deep learning; neural networks.