Machine Translation
- Type
- NLP text-to-text generation task
What does it do?
Machine translation (MT) is a natural-language-processing task that automatically converts text in one language into another.
Modern MT is built on neural sequence-to-sequence models — the attention mechanism introduced by Bahdanau et al. (2015) and the transformer architecture of Vaswani et al. (2017) — and it widens access to sources for multilingual data analysis and research.
The intuition behind the method
Reading material in a language you do not speak blocks you from using it. Machine translation removes that barrier: instead of translating word by word, a model reads the whole source sentence, builds an internal representation of its meaning, and then writes out an equivalent sentence in the target language, attending to the most relevant source words as it produces each output word.
History
Machine translation moved from rule-based and statistical systems toward neural models. Bahdanau et al. (2015) introduced an attention mechanism that lets the model jointly learn to align and translate, removing the bottleneck of compressing a sentence into one fixed vector. Vaswani et al. (2017) then proposed the transformer, which relies entirely on attention and became the foundation of modern translation systems.
When to use it
Use machine translation when you have source-language text and need it in another language, and a trained model for that language pair is available. It suits widening source access in multilingual data analysis and research. For specialised-domain text, expect to fine-tune the model. At least a small amount of text is needed to translate.
Limitations and considerations
- Requires an available trained model for the specific source and target language pair.
- Specialised-domain translation may need fine-tuning before it is accurate enough.
- Output must still be checked for fidelity and fluency; errors can be subtle.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Widens access to sources by making text in other languages usable for analysis and research.
- Neural attention and transformer models translate whole sentences in context rather than word by word.
- Scales from short passages to large multilingual corpora through the same model.
Applications
- Widening access to non-native-language sources in multilingual research.
- Translating documents and corpora for cross-language data analysis.
- Preparing foreign-language text for downstream NLP tasks.
- Supporting fine-tuned translation for specialised domains such as legal, medical, or technical text.
How it works
How it works
Source text is tokenised and prepared so the model can read it. A trained model for the source and target language pair must be available; for specialised-domain translation, fine-tuning the model on in-domain text may be needed before it is used.
The model reads the source sentence and builds an internal representation of its meaning. The attention mechanism lets it learn which source positions to align with as it works, rather than compressing everything into a single fixed summary.
The target-language sentence is generated one unit at a time. As it produces each output word, the model attends to the most relevant parts of the source, jointly learning to align and translate so that meaning is preserved across languages.
The translated output is checked for fidelity to the source and fluency in the target language. Domain-specific errors are inspected, and fine-tuning is reconsidered when the general model is not accurate enough for the material at hand.
Common pitfalls
- Applying a general model to specialised-domain text without fine-tuning yields inaccurate translations.
- Using a model that does not cover the required language pair gives unusable output.
- Trusting raw output without checking fidelity and fluency can propagate translation errors into downstream analysis.
Frequently asked questions
What kind of model does machine translation use?
Modern MT uses neural sequence-to-sequence models. The attention mechanism (Bahdanau et al., 2015) lets the model align source and target as it translates, and the transformer architecture (Vaswani et al., 2017), built entirely on attention, underpins current systems.
Do I need a model for my specific language pair?
Yes. A trained model for the source and target language pair must be available. Without coverage for that pair, the system cannot produce usable translations.
Why might translation quality be poor on my texts?
Specialised-domain material often needs fine-tuning. A general model may translate everyday language well but struggle with domain-specific terminology until it is adapted on in-domain text.
Should I trust the output directly?
Check it. Translated output should be evaluated for fidelity to the source and fluency in the target language before it feeds into further analysis, since errors can be subtle.
Related methods
Sources
Bahdanau, D., Cho, K. & Bengio, Y. (2015). Neural Machine Translation by Jointly Learning to Align and Translate. International Conference on Learning Representations (ICLR).
Open publication referenceCopied.Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L. & Polosukhin, I. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems (NeurIPS).
Open publication referenceCopied.