CSCI-GA.2590 - Natural Language Processing - Spring 2013 -- Prof. Grishman

Lecture 13 Outline

April 30, 2013

Machine Translation [J&M Chap 25]

Objectives

Terminology:  source language and target language

Language Differences:  Obstacles to good MT [J&M 25.1]

Classical MT [J&M 25.2]

Machine Translation was the first focus of natural language processing. Systems were developed starting in the late 1950's, and some degree of success was achieved in translating between closely related European languages. It also led to an understanding that FAHQT [fully-automatic high-quality translation] would require a much deeper understanding of language [ALPAC report].

Statistical MT [J&M 25.3]

Language Model

A Very Simple Translation Model

Putting it Together: the Decoder

Once the translation model and language model have been trained, translation can be performed by evaluating

argmaxE P(F | E) P(E)

This is a search for the most likely sentence E. Clearly we cannot search through all English sentences; not even through all sentences composed of words which translate into some word in F. To make the search more efficient we

Better Translation Models