Statistical-based machine translation for prepositional phrase using Link Grammar

The Internet provides most information in English language. However, most people are non-English speaker. Hence, it will be a great advantage for the people to have reliable machine translation. The expectation from the Machine Translation (MT) is sentences translation in a natural way based on the contexts. There are some reasons of why MT is hard to develop. One of the cause is there exists many mapping from source to target words in one-to-many, many-to-one, and many to-many. For example, an English preposition can have more than one translation in Indonesian words depend on the context. This research tries to find a technique that can provide a suitable translation of the English prepositions to the Indonesian words. The technique will use LG formalism and statistical method. The meaning of preposition in a sentence depends on the preceding word and the following word. The word before the preposition is a word explained by the preposition. Meanwhile, the word after a preposition is a word that explains the preposition. Knowing that both preceding and following words are not always precisely in the right and left positions of the preposition, we then need a method to find those connected words. We try to handle the problem by using LG formalism. LG formalism consists of a set of words and each word has a linking requirement. Using this linking requirement, the words that have a link to the preposition can be resolved. After obtaining the connected words, the translation is done using statistical method. Statistical method translates the preposition based on the probability of the relation which is obtained from LG formalism. This work expresses its relation in the bigram form. If an English preposition bigram has more than one Indonesian translation, then the output is the bigram that have the biggest probability. In this research, the combination between LG formalism and statistical method give a better English preposition translation than the existing methods.