The 3 most common types of machine translation are rule-based, statistical, and neural machine translation.
Rule-based machines translation relies on a predefined set of linguistic rules that helped the software transfer text between languages. It had low quality outcomes, and it required a significant amount of human post-editing. Rule-based machines translation is hardly used today.
Statistical machine translation builds a statistical model of the relationships between words, phrases, and sentences in a given text. It applies the model to a second language to convert those elements to the new language. Thereby, it improves on rule-based translation but shares many of the same issues.
Neural machine translation uses AI to “learn” languages. It uses neural networks to replicate the human translation process at an accelerated pace. These networks are made up of layers of connected nodes that transmit signals, much like the networks of neurons in the human brain do. As opposed to running a set of predefined rules, a neural network is responsible for encoding and decoding the source text — one layer encodes the source text and another decodes it into the target language, using knowledge acquired from training on massive datasets. Then, the third component of neural network, which “connects the dots” between the encoder and decoder, the so-called “attention mechanism”, which helps the decoder accurately weigh how much each word in the input sentence should impact the translation of the current output word.
We’ll discuss the pros and cons of neural machine translation next time.