Google Translate
Google's translation service, which switched from phrase-based statistical translation to neural sequence-to-sequence models in 2016 and later to Transformers.
Also machine translation · GNMT · Google Neural Machine Translation · neural machine translation
Google, launched 2006; neural system from November 2016
It used to translate sentences by gluing together bits it had seen translated before. Now it reads the whole sentence, understands how the words relate, and writes a fresh translation.
Where it fits
+All relationships as a list
Is a kind of / part of
Siblings
- AI Agents and Tool Use
- BERT
- Direct Preference Optimization
- Embedding
- Fine-Tuning
- GPT
- Hallucination
- In-Context Learning
- Keyboard Autocomplete
- Large Language Model
- LLM Evaluation and Benchmarks
- Perplexity
- Pretraining
- Prompting
- Retrieval-Augmented Generation
- RLHF
- Sampling and Temperature
- Scaling Laws
- Tokenization
- Vector Database
- word2vec
How it works
Google Translate launched in 2006 with statistical machine translation: it learned from millions of existing translations which phrases tend to correspond, then pieced translations together phrase by phrase. The results were often understandable but clumsy.
In November 2016 Google switched many languages to Google Neural Machine Translation (GNMT):
- A deep LSTM encoder reads the whole source sentence.
- A decoder writes the translation word by word, using attention to focus on the relevant source words at each step.
This is the sequence-to-sequence architecture, and the jump in fluency was large enough to make headlines. A follow-up showed zero-shot translation: a single multilingual model trained on English to Japanese and English to Korean could translate Japanese to Korean directly, hinting that it had learned a shared representation of meaning.
The Transformer, introduced by Google researchers in 2017 precisely for translation, later replaced the LSTMs. Today general LLMs are also strong translators, and the boundary between “translation app” and “assistant” is blurring.
Known weaknesses include low-resource languages and bias: translating from languages without gendered pronouns, systems have defaulted to “he is a doctor” and “she is a nurse” (bias and fairness).
Everyday use cases
- Case 01
How you've probably met it
Translating a menu with your phone camera, a foreign web page in your browser, or a conversation in real time while traveling.
Strengths, weaknesses and when to use it
Strengths
- Covers a very large number of languages, for free, in text, speech and images.
Weaknesses
- Errors with rare languages, idioms and context; can reproduce biases such as gendered defaults for professions.
When to use itAs the everyday example of sequence-to-sequence learning and the move from statistical to neural NLP.