DOI: 10.5176/2251-3566_L317.5
Authors: P.P.Giridhar
Abstract: This intervention privileges RBMT over other MT’s because other species of MT leave a lot to be desired, don’t hit the bull’s eye, don’t touch the bottom line. They are not underpinned theoretically soundly. The argument is that grammars of human languages are not stochastic ones, that there is no doing any MT without arriving at (nonstochastic) rules that govern natural language, and going by the quality of the output of corpus-based statistical approaches, post-edited crowd- sourced approaches at least in India, one could say that the machine can’t discover rules of language the way humans can. Or can it? These approaches lead you to a conception of natural language that is not right: trial- and-error, hit-or-miss, stochastic, not very sharp-edged, grammatical by stages, in fits and starts, in bits and pieces. Reality checks of the MT scene in India are in fact not reality checks in as much they fail to tell you about the quality of the output. The one available on the Net titled ‘MT Set for a Quantum Leap in India’ is delightfully bereft of conviction. (Our earnest and rationally grounded hope is that MT systems like METEO the Canadian MT system for weather reports since 1977, Systran, Moses, Asia Online etc. are rigorously rule-based and their output is (consequently) nonstochastically grammatical.) Approaches to (human quality) MT other than RBMT can only subserve RBMT. About 85{6e6090cdd558c53a8bc18225ef4499fead9160abd3419ad4f137e902b483c465}-90{6e6090cdd558c53a8bc18225ef4499fead9160abd3419ad4f137e902b483c465} accuracy is pretty much doable for any language pair in three-four years of smart informed team-work. The article also seeks an answer to why 100{6e6090cdd558c53a8bc18225ef4499fead9160abd3419ad4f137e902b483c465} accuracy is not possible, at least as of now. The essay however is more an object in the world of ideas than an implementation on the ground.
Keywords: MT, rule-governed, crowd-sourcing, user-validation, postedited
Updating...