Data Pre-Processing to Train a Better Lithuanian-English MT System

Pried -as ir Protokol -as yr -a neatskiriam -a ši -o Susitar -imo dal -is. Prefixes separated, endings replaced by tense and number feature values System #2 Prefixes separated, all endings replaced by number feature values and verb endings also by time feature values System #3 Prefixes separated, endings deleted System #4 As Lithuanian is highly inflected language, the words change the form according to grammatical function. That means that the endings of nouns, pronouns, adjectives, numerals and verbs change depending on certain features. English instead does not have such a rich feature system. This difference between languages significantly impacts word and phrase alignment when training an SMT system. Typically one or two forms of an English word have to be aligned to more than ten different surface forms of a corresponding Lithuanian word. Lithuanian verbs have prefixes indicating negation and other semantic features while English verbs do not have prefixes and such information is expressed using modifying words. Many word forms are not as common as others in the corpus, therefore a Lithuanian-English SMT system does not translate all word forms equally well. It is very common to get many out of vocabulary words when translating from Lithuanian into English. Chosen approach