Types of machine translation

Machine translation technology has come a long way since its early days. Over time, several approaches have emerged. The most popular of them include:

• Rule-based machine translation (RBMT) - the first commercially used machine translation systems that came out in the 1960s. These models rely on the idea that every language has a large set of grammatical, syntactic and semantic rules. Linguists define the rules for sentence structure, word order and phrasing in both the source and target languages. The system then generates sentences by matching source language words to dictionary entries. RBMT isn't used today because of its drawbacks, such as: long development and poor translation quality.
• Statistical machine translation (SMT) - systems that are based on statistical probabilities. Statistical models appeared in the 1990s and were trained by analyzing large amounts of bilingual texts in the source and target languages to learn associations between words. To produce translations, SMT finds equivalent expressions and defines how other similarly constructed texts should be translated. Statistical approach offers good translation quality, requires less post-editing and can handle ambiguity better than RBMT.
• Neural Machine Translation (NMT) - more advanced, probabilistic models that rely on artificial neural networks, particularly deep learning architectures. NMT uses sentence context to generate the best translation option. It maps a sequence of words from the source language to its equivalent sequence in the target language. Then, it compares the generated output with the expected translation and makes necessary adjustments until the translation is grammatically correct and sounds “human-like”. Neural networks require a lot of data to learn and perform their tasks successfully. However, once fully trained, they continuously improve the quality of the output without any human intervention. Various translation platforms use NMT as their model, such as DeepL, Google, Systran and Yandex.

• Large-scale language models (LLM) - very large deep learning models that are trained on vast amounts of data to grasp language patterns and relationships between words. They can generate language from scratch and be more flexible when interpreting text or prompts. LLM started to become popular around 2018 when the ChatGPT model GPT-1 was introduced. The first version wasn't trained specifically for translation, which became possible in other models such as: GPT-2, GPT-3, GPT-4 or BERT. LLM bring significant progress to machine translation technology, mainly because they can translate with context. They analyze the text and try to understand its meaning before generating the output.