Divulgação - Defesa Nº 208
Aluno: Arthur Flor de Sousa Neto
Título: “Towards the Natural Language Processing as Spelling Correction for Offline Handwritten Text Recognition Systems”.
Orientador: Prof. Byron Leite Dantas Bezerra
Coorientador: Alejandro Héctor Toselli Rossi
Data-hora: 30/Julho/2020 (14:00h)
Local: Escola Politécnica de Pernambuco – Formato Remoto (https://meet.google.com/oof-gaze-gde)
Resumo:
“The growing demand for portability of physical manuscripts to the digital medium makes the use of more robust and automatic mechanisms common in offline Handwritten Text Recognition(HTR) systems. However, the great diversity of application scenarios and writing variations, bring challenges to the text recognition precision and, to minimize this problem, the optical model can be used in conjunction with the language model, in which it assists in decoding text through predefined linguistic knowledge. Thus, in order to improve the results, character and word dictionaries are created from the dataset used, causing the linguistic restriction within the HTR system. In this way, this work proposes the use of spelling correction techniques for text post-processing in order to obtain better results in the final stage and eliminate the linguistic dependence between the optical model and the decoding step. In addition, an encoder-decoder neural network architecture and training methodology are also developed and presented to achieve this goal. To validate the efficiency of this new approach, we conducted an experiment using: (i) five datasets of lines of text already well known in the HTR field, including a set that corresponds to a combination of all of them (All in One); (ii) three state-of-the-art optical models; and (iii) eight spelling correction techniques within the field of Natural Language Processing, varying between traditional statistical and more recent approaches, such as neural networks. In this way, the results of the techniques combinations are presented and discussed in each dataset individually. Finally, the proposed spelling correction model with the best performance is analyzed statistically, through the metrics of an HTR system and considering all the results obtained from the combinations, reaching an average sentence correction of 65%. This means a 54% improvement over the traditional method of decoding on tested datasets. In addition, other simpler statistical techniques are also discussed, bringing relevant results in some applied scenarios.”
O Programa
