Abstract:
This article presents the results of the formation, training and
performance evaluation of models with the Encoder-Decoder and Sequence-To-Sequence (Seq2Seq) architectures for solving the problem of supplementing
incomplete texts. Problems of this type often arise when restoring the contents
of documents from their low-quality images. The studies conducted in the work
are aimed at solving the practical problem of forming electronic copies of scanned
documents of the «Roskadastr» PLC, the recognition of which is difficult or
impossible with standard means.
The formation and study of models was carried out in Python using the
high-level API of the Keras package. A dataset consisting of several thousand
pairs was formed for the purpose of training and studying the models. Each pair
in this set represented an incomplete and corresponding full text. To evaluate the
quality of the models, the values of the loss function and the accuracy, BLEU and
ROUGE-L metrics were calculated. Loss and accuracy made it possible to
evaluate the effectiveness of the models at the level of predicting individual
words. The BLEU and ROUGE-L metrics were used to evaluate the similarity
between the full and reconstructed texts. The results showed that both the
Encoder-Decoder and Seq2Seq models cope with the task of reconstructing text
sequences from their fixed set, but the Seq2Seq transformer-based model achieves
better results in terms of training speed and quality
Key words and phrases:deep learning models, encoder-decoder, sequence-to-sequence transformer, text recovering, BLEU, ROUGE-L, Keras, Python.
Fulltext:
PDF file (2364 kB)