Reordering Rules for Machine Translation Using Least General Generalizations

Machine translation has been one of the key problems in computational linguistics. An important part of the same is to get the order of words in the translated sentence correct. Numerous approaches exist that range from treating sentences as plain strings to those relying on sentence structure. Among the latter, manually written rules as well as rules learnt by machine have been known to benefit the intended purpose.

In this talk, we discuss our work that learns rules for ordering words in the translated sentence using structural information about the original sentence. We view the problem as that of learning a declarative set of instructions and thereby deploy a technique used in program synthesis that uses an operator called LGG (Least General Generalization) (this work was done jointly with Ganesh Ramakrishnan and Amitabha Sanyal at IIT, Bombay).