Improving Machine Translation Output with Lightweight Preprocessing and CNN-Based Quality Estimation

Abstract

Ensuring high-quality output in Neural Machine Translation (NMT) systems remains a central challenge, especially in applications with critical fluency, grammatical accuracy, and semantic fidelity. While significant advancements have been made in model architecture, less attention has been given to the role of data preprocessing and post-translation evaluation, both of which are essential for enhancing translation reliability and scalability. This study introduces a dual-stage framework that integrates linguistic preprocessing techniques and a Convolutional Neural Network (CNN)-based classifier for translation quality assessment. We use the English--Spanish Translation Dataset by Lonnie Qin to train a lightweight Seq2Seq NMT model with and without preprocessing steps, including tokenization, lowercasing, lemmatization, and normalization. Translation quality is evaluated using BLEU, METEOR, and ROUGE metrics. A 1D CNN is trained as a lightweight post-translation screening model using BLEU-derived weak labels for large-scale binary supervision and a smaller human-labeled subset for finer-grained validation. The purpose of this classifier is not merely to reproduce a fixed BLEU cutoff, but to learn reusable quality patterns that can support rapid quality flagging in settings where repeated reference-based evaluation or manual review is impractical. It is important to note that the CNN is not used as the translation generator itself, but as a lightweight post-translation quality estimation module, selected to examine whether reliable sentence-level quality screening can be achieved in resource-conscious and real-time settings without resorting to heavier transformer-based classifiers. Experimental results demonstrate that preprocessing consistently improves the translation accuracy of this lightweight model, with BLEU scores increasing by 8.4 points and METEOR by 4 points; these gains are specific to the simple Seq2Seq model studied here and are not claimed to transfer to stronger architectures. The CNN classifier achieves 88.7 accuracy in binary classification and a macro F1-score of 0.82 in multi-class evaluation. The integrated pipeline improves both the generation and validation of translations, making it suitable for real-time quality assurance and educational use cases. The proposed approach highlights the often-underestimated impact of preprocessing and the efficiency of CNNs in evaluating translation quality. Together, they form a robust, scalable, and adaptable solution for improving translation outputs. This framework offers substantial potential for deployment in educational and professional language-support settings, while domain-specific applications such as legal or medical translation require further validation on specialized corpora.