Capabilities
Machine TranslationMT
Also known as: AI translation
AI that translates text (or speech) from one language to another; one of the oldest goals of AI and now remarkably fluent.
Machine translation is AI that converts text from one language into another, the technology behind translation apps and websites. It has become impressively fluent, handling everyday text well across many languages. It is not perfect, though: tone, idioms, names and specialised or legal wording can still trip it up, so important translations deserve a human check.
Machine translation (MT) automatically translates between languages. It was one of the founding challenges of NLP and has improved dramatically with deep learning.
How it evolved
- Rule-based → statistical → neural. Neural MT, and especially the transformer (which was originally introduced for translation), produced the big leap in fluency.
- Today, general LLMs translate well as one of many skills, often without being purpose-built for it.
Where it still struggles
Idioms, cultural nuance, ambiguous context, rare languages, and domain-specific terminology (legal, medical). Quality also varies a lot between high-resource languages (lots of training data) and low-resource ones.
Neural machine translation models the conditional probability of a target sequence given a source, historically with encoder-decoder RNNs and attention, then decisively with the transformer, whose architecture was introduced for MT (Vaswani et al. 2017). General LLMs now perform strong zero-/few-shot translation, and the task increasingly folds into broader multilingual and multimodal systems (including speech-to-speech).
Enduring issues: disparity between high- and low-resource languages (data imbalance), faithfulness on named entities, numbers and terminology, discourse-level coherence and pronoun/gender resolution, and domain adaptation. Tokenizer fragmentation also disadvantages some scripts in cost and quality. Evaluation moved from n-gram overlap (BLEU) toward learned metrics (COMET) and human assessment, which better capture adequacy and fluency. Despite near-human quality on many high-resource pairs, reliable use in high-stakes settings still warrants human review.