HOW MACHINES TRANSLATE
First, I want to point out how conventional machine translation systems work. The course of human translation which consists of reading and comprehending a text, transfering words and converting grammar structures, and finally, if necessary, improving the translation, is a great contrast to most translation softwares' operating principles which works like the following scheme, explanation and a simple example demonstrate:
First of all parsing, a preparation which splits a text, leads the way. Documents are divided into sentences, sentences are partitioned into single words, terms or formatting information. Time or date information, names and terms of compound nouns, for example, must not be torn apart.
2. Morphological Decomposition
Bilingual dictionaries consist only of word stems and words' basic forms because it is impossible to save each form and modification of words in both the source and target language. To illustrate the problem, e.g. one German verb would require six entries in the dictionary only for its present indicative forms. So if each form were added, the size would explode. That's why an additional step to identify the basic forms and stems of the words – especially verbs and compound nouns – is necessary before the transfer can start. Often these algorithms are called "morphological decomposition".
A translation's quality is determined by both syntactic and semantic analysis. Translation systems must be able to deal with sentential and grammatical structures, detect relations, references and general contexts to generate useful translations.
After a text has been analysed, the systems try to find equivalents to the individual parts of a sentence by the aid of dictionary look-ups. The preparations from step one and two make these look-ups possible. The transfer is also influenced by step three because it considers the previous analysis' results. Some machine translation systems use so-called "intelligent dictionaries". These help to distinguish between the meanings of words which have more than one corresponding translation in the target language with regard on the subject of the whole text.
Finally the new sentences in the target language are changed with respect of structures and grammatical constructions. If it is necessary, the system changes the position of a word, for example.