Automatically Filtering Domain Specific Data from Web Mined Parallel Corpora (Progress Review 4)

Domain adaptation remains a critical challenge in Neural Machine Translation (NMT), particularly for low-resource languages where high-quality domain-specific parallel corpora are limited. Large web-mined parallel corpora are often noisy and contain mixed-domain content, requiring effective filtering and selection mechanisms to be useful for domain-specific training. This research proposes an automated pipeline for constructing domain-specific training datasets from web-mined parallel corpora without human intervention. The study evaluates multiple domain relevance scoring techniques, including similarity-based, divergence-based, classifier-based, and Domain-Specific Importance Resampling (DSIR)-based methods. Experimental results demonstrate that Discriminative DSIR-based approaches consistently achieve strong translation performance across multiple domains and language pairs. Overall, this work presents a scalable automated pipeline for generating high-quality domain-specific datasets, improving translation performance in low-resource machine translation settings.