For the last few years, one of the biggest terms in the translation industry has been neural machine translation. While it is often used, to many people, the term is somewhat ambiguous and confusing. Especially when it comes to the first word, neural: though machine translation is obviously translation done by a machine, what about the process that makes it neural?
Before answering that, let’s examine the word itself. The word neural, meaning ‘pertaining to a nerve or nerves’ originally comes from the Greek word for nerve, neuron.
First used in T.R. Jones’ contribution to Robert Bentley Todd’s The Cyclopaedia of Anatomy & Physiology (1847), stating: “The caudal vertebra of the Fish..have the neurapophyses and neural spine as well as the hæmapophyses and hæmal spine.”, the word was initially meant to denote the side of the body where the spinal cord is located.
Soon used to encompass the physiological activity of the entire nervous system, our term was first used to relate to something specifically outside of the nervous system (but relating to the nature of it) in William H. Dallinger’s 1882 work, The Creator, writing that the “Mind is inseverably associated with neural matter.”
Relative to technology, neural first appeared in The Quarterly Review of Biology (Vol. 22, 85/2) in 1947, which wrote that: “The topics considered, which include..the general characteristics of neural networks, are of undeniable importance.”
The terms neural computer and neural computing were both first recorded in use back in 1986, with a paper introducing the Electronic Neural Computing. And neural computing as a corporate advantage, along with the relation between its learning curve and precision and Big Data – in the 1993 February issue of Computer Weekly: “Neural computing, which mimics the behaviour of the brain, is well suited to identifying patterns and analysing trends in the large amount of data which organisations routinely collect about their production processes, operations, markets and clients.”
Although we all know what statistical machine translation is – where text is translated by a machine that uses statistical analysis to determine what the individual words likely mean in the context of the surrounding words – many people don’t understand what makes neural machine translation, well, neural. The use of the word comes from the program’s ability to learn like a biological neural network. Much in the way that children (and some adults, admittedly) learn the painful lesson of why you don’t touch a hot frying pan, neural machine translation possesses the neural ability to learn, from training and usage, how you mean the words that you use based within the context of the entire sentence. Given, it can still make mistakes, as the technology is not yet expansive enough to take entire works/documents into consideration, but, by mimicking the ability to learn usage, it represents a great leap from traditional, statistical machine translation.
