What is Neural Transfer?

Neural transfer is a method which, like the human brain, uses associations for selecting translation alternatives.

It helps in choosing the correct translation if there are several possible translations available.
(Example: Gericht = court and Gericht = dish. Or: bill = Gesetz and bill = Rechnung)

The central idea is: What do humans do in such cases to find the correct translation?
They analyze the conceptual context of the word in question.
If they find words like "Anwalt", "Justiz", "Richter", etc., used in the context of "Gericht", they think: in this context it must be translated as "court". If, on the other hand, they find words such as "Nudeln", "Salz", "Gewürze", etc., they conclude: in this context it must be translated as "dish".

The next question is: How do humans come to this conclusion?
By knowing that these things are related to each other (this is the concept of "world knowledge"). Such world knowledge is stored and networked in the human brain (forming a "neural network").

The neural transfer function tries to replicate this associative human approach. Using linguistic and neuroinformatics methods, enormous quantities of text are analyzed to identify which concepts are usually used in context with each other. With more than 1.5 billion words the Linguatec corpus is the largest collection we know! Printed and piled up it would be a stack of paper more than 125 meters high.

These terms are extracted and saved in an associative memory (a neural network). The information held here would indicate, for example, that "plant" is probably translated as "Pflanze" if used in the context of "flower", "water", etc., but is probably translated as "Werk" if used in the context of "electrical", "chemical", "workforce", etc.

This neural network is activated if the system encounters a term with several meanings:

It looks at the context in the text,
compares it with its knowledge of such contexts as saved in the neural network, and
decides on a translation.

This makes it possible to find translation alternatives which have so far not been possible to identify in machine translations because previous systems a) only analyze the individual sentence (instead of the entire context), and b) do not have this associative knowledge capability.

Of course, no system is perfect and the intelligent human can always find cases in which the ignorant machine chooses the wrong option, for example, if the "parliament" "pays" its "phone" "bill".