The Dream
The dream started tens of years ago to make a machine that can automatically translate from one human language to the other. Today, the dream still holds its old name "machine translation" but has now become one of the applications of artificial intelligence, a branch of computer science that tries to make the computer solve problems using different methods than the ones classically used when programming a computer.
Many machine translation systems have been built today but, unlike in chess where IBM's Deep Blue was able to beat Kasparov the world's chess champion, non of the machine translation systems present today are able to reach the abilities of a a professional human translator. The reason behind this is that translation from one natural language to the other is a complex process that involves complex types of knowledge, reasoning and common sense. Advancements in natural language processing, a branch of artificial intelligence, have not yet led to the development of a machine translation system comparable in skill to that of human translators.
Gist Translation
Today, there are basically two applications of machine translation in use. MT is used in many web sites on the Internet to provide you with the gist of a text written in a language you do not understand. You provide it with the address of a web page or with some text in one language and it gives you the translation in the another language of your choice. Usually the translation is poor specially with text that uses long sentences and also text that is non-technical. Sometimes the translation produced sounds funny and many times it gives you wrong translations of sentences. Moreover, reading such machine translated text could be frustrating, while your brian is trying to make extra effort figuring out what the real meaning is behind such badly translated text. That said, gist translation can help you determine the general idea behind a web page and provide you with the main points in it. It can also give you a better translation if the text you are translating is technical or written in very simple and short sentences that use simple grammatical constructs and exceptionally clear language.
Restricted Translation
The other application of MT also used today is in translating product manuals. Some large international companies have made use of machine translation to shorten the time to market their products in different countries of the world by shortening the time needed for translating its manuals. Not only have they managed to cut translation time but also price as it costs a lot to have human translators translate the whole content of your product manuals say to 22 different languages, while it can cost a fraction of that price and a fraction of the time needed to translate the same thing by computers using machine translation. Unlike in gist translation, machine translation software used for translating product manuals can produce much higher quality. The reason behind this is that the company takes care to write the original manual (say in English or in German) using very simple grammatical rules and a restricted vocabulary. This makes it easy for specially designed machine translation software to 'chew' on those simple sentences and produce their equivalents in other languages. Moreover, the original manual is pre-edited before it is fed into the machine translation software in order to make sure it is easy for it to translate. Not only that, but after the translation is done, it is post-edited by human editors and checked for errors in order to come out with a good translation at the end. The use of restricted simple grammar, controlled vocabulary, pre-editing and post-editing results in a high quality translation of the original product manual at the end. Though this process may take some time, yet it is sure much less time consuming and more cost effective than paying for human translators to translate the manual into all those languages required.
Machine Translation Methods
Researchers have attempted, and still are, different methods to tackle the problem of machine translation. One popular method is to try and transform the source sentence to the target sentence using a transformational model that 'maps' the structure,g rammar, meaning and other aspects of one language to the other. This model is widely used today and is good for systems that translate between only one pair of languages. Another method is to convert the source sentence into an abstract representation of its meaning then convert this abstract representation to a sentence in the target language according to the language model of the target language. This method works well for systems that attempt to translate between many sets of different languages. In this case, a single model is create for each of those languages, while in the transformational method mentioned earlier, a transformation model will have to be build between each pair of languages, which will lead to an exponentially larger number of models as the number of language pairs supported increases. One of the new methods of MT is example based machine translation. This method uses neural networks to attempt to learn how to translate from already translated sentence pairs. Not much success has yet been made using this method but a lot of research is being made in this direction and it does look promising specially for language pairs that have a lot of bilingual material already translated in order to feed the system with to learn from.