Machine Translation

Currently, VISL has 3 projects within the area of Machine Translation (MT), all of which are real life projects that address the full variety of natural language text, and are not in any way restricted to a sample lexicon or a limited number of sentence structures.

In all three projects, our MT strategy relies on a detailed and robust grammatical analysis of the source language (SL) side. The systems' parsers are based on Constraint Grammar technology (PALAVRAS and DanGram, respectively), and produce word based morphological and functional information as well as syntactic tree structures. Semantic disambiguation uses semantic prototype discrimination and valency instantiation. In addition, specific translations or translation patterns are mapped with context dependent CG-rules. The systems differ, however, in their approach to syntactic transfer. Po-Da and Da-Eo use syntactic transformation rules based on tag sequences, while Da-En uses movement rules based on full dependency trees.

The objective of VISL's machine translation is not "perfect translation" (which - if at all possible - would involve a great deal of artificial intelligence and knowledge about the world), but rather to provide a target language (TL) version of the input text which, though in places somewhat ungrammatical, lexically unorthodox or even incomplete, will allow monolingual SL native speakers to understand and sensibly post-edit the text in question.