Optimal Translation Boundaries for BTG-Based Decoding

Abstract

This paper proposes a method for inducing translation boundaries as soft constraints for Bracketing Transduction Grammar based (BTG-based) decoding. Translation boundaries used in previous research are extracted from left-most synchronous trees generated by a deterministic algorithm. Translation boundaries in this research are extracted from induced synchronous trees, which are statistically optimal and more balanced than the left-most synchronous trees. Experiments show that induced translation boundaries are more consistent than those extracted from left-most synchronous trees, resulting in significantly better performances over the strong baseline.