Predicting ESL Translation Errors with Hybrid Hybridization of Deep Neural Networks and Decision Trees
Abstract
MT and Second-Language GEC have not yet been fully integrated, despite their shared goal of producing accurate and fluent output. The lack of a universally accepted standard and the shortcomings of existing automatic metrics make it difficult to analyse ESL translation problems in machine-translated texts. For example, BLEU and METEOR provide only single-score evaluations without any insight into where the errors originated. Manual inspection is more accurate, but it takes more time and there aren't any standards on how to classify errors. Our innovative DTCN model for hierarchical error detection in ESL translations is here to fill this need. There are three parts to the model: training, feature extraction, and preprocessing with data normalization. We were able to extract more discriminative information using our feature extraction method than with more conventional methods like principal component analysis. The DTCN recreates CNN output layers using tree-based hierarchies, applying the structural advantages of decision trees and the representational capability of CNNs. The experimental findings show that the suggested model outperforms baseline models by a wide margin, reaching an amazing 99.21% accuracy. In addition to improving error identification in ESL translation, this method encourages more cooperation between MT and GEC studies.