Dynamic contraction and expansion of heuristic beam width based on predicted probabilities

Abstract

A method, computer program product, and computer system for translating, using a beam search, a source sentence in a source language into a target sentence in a target language by an iterative process. Each iteration of the iterative process includes: generating, using a sequence-to-sequence model, probability vectors of conditional probabilities of respective vocabulary words in the target language being translations of a source word in the source sentence; sorting the probabilities in the probability vectors; generating probability difference vectors containing numerical differences between adjacent elements in respective sorted probability; vectors determining, by a fully connected neural network (FCNN), a best beam width B using the probability difference vectors as input to the FCNN; selecting B vocabulary words and B target vectors corresponding to the B highest conditional probabilities; and after all words in the source sentence have been translated, outputting the B target vectors generated in the last iteration.