Desarrollo y aplicación de un modelo computacional para el procesamiento de lenguaje natural : detección de errores ortográficos en textos obtenidos por reconocimiento de caracteres ópticos (OCR) mediante Representaciones de Codificador Bidireccional de Transformers (BERT).

Abstract

This work focuses on natural language processing (NLP) tasks, specifically addressing the processing and alignment of a parallel corpus by identifying the necessary tags for applications in the development of a supervised model for detecting spelling errors obtained through an OCT tool.

To achieve this goal, the Fine-Tuning technique is applied to leverage the encoder of a pre-trained Transformer architecture model, known as Bidirectional Encoder of Transformers (BERT). In this manner, the model is utilized to process information and contextualize the received data, forming a specific architecture for the decoder. The decoder is constructed with the required layers to structure the classification layer, which aims to assign classes accordingly to words, categorizing then as either correct or incorrect.