Improved BTG-based Preordering for SMT via Parallel Parameter Averaging: An Empirical Study

Abstract

Preordering has proven useful in improving the translation quality of statistical machine translation (SMT), especially for language pairs with different syntax. The top-down bracketing transduction grammar (BTG)-based preordering method (Nakagawa 2015) has achieved a state-of-the-art performance since it relies on aligned parallel text only and deos not require any linguistic annotations. Although this online learning algorithm adopted is efficient and effective, it is very susceptible to alignment errors. In a production environment, in particular, such a preorderer is commonly trained on noisy word alignments obtained using an automatic word aligner, resulting in a worse performance compared to those trained on manually annotated datasets. In order to achieve better preordering using automatically aligned datasets, this paper seeks to improve the top-down BTG-based preordering method using various parameter mixing techniques to increase the accuracy of the preorderer and speed up training via parallelisation. The parameters mixing methods and the original online training method (Nakagawa 2015) were empirically compared, and the experimental results show that such parallel parameter averaging methods can dramatically reduce the training time and improve the quality of preordering.