Parallel corpus
A parallel corpus is a large collection of text aligned with its translation, sentence by sentence, across two languages. It is the raw material machine translation systems learn from: the more, and the better aligned, the parallel data for a language pair, the stronger the engine that can be trained on it. Translation memories are a form of parallel corpus.
How it works
Parallel data is gathered from already-translated sources, official documents, websites, subtitles, then cleaned and aligned so each source segment sits beside its translation. Training reads these pairs and learns the mapping between the languages.
Quality and quantity both count. Noisy or misaligned pairs teach the engine bad habits, and thin data leaves it guessing, which is why high-resource pairs translate so much better than low-resource ones.
How SourceTarget uses it
The translation memories SourceTarget builds from a client's work are a form of parallel corpus, and they stay the client's. Clean, well-aligned parallel data is what improves consistency and, where volumes justify it, what a custom engine is trained on.