Abstract

Introduction l"Jxalnl)le-lgase(l i4aehine Translation relies on a, collection oF textual units (usually sentences) a.n(I their 1,ra. nsla.tions. New text I)e tra. nsla. ted is ma. Cched a.ga.inst the sourceh ngua,ge ha. lf of' the eollecLk)n, and the covve- spentling tva.nsla.tions from the ta.rgeC-la. ngua.ge half a. re used to generate a. tra. nsltion of new text. 13xperience with severa.1 language pairs has shown that producing a.n EI1MT system which 1)verities rea.sona. ble [va.nslat, ion coverage of resCricCe(t texCs using simple texCual matching requires on the or(let of two million words of' l)Ve-translated texts (one million wovcts in each hngua.ge); iF either la. ngua.ge is highly inflecting, l)olysynthetic, or (worse yet) a.gglu Lina.tive, even move text xvill 1)e required. It may I)e difficult, tine-eonsumilg, ;red expensive to obtain tha.t much pa, ra,llel text, pa. vtieula. rly f'or lesser-used la. nguage pa. irs. 'l'hus, if one wishes to develop a, now translator ra. pi

Citations