Morphology In Statistical Machine Translation From English To Highly Inflectional Language

In this paper, we investigate the role of morphology in phrase-based statistical machine translation (SMT) from English to the highly inflectional Slovenian language. Translation to an inflectional language is a challenging task because of its morphological complexity. Rich morphology increases data sparsity and worsens the quality of statistical machine translation. The idea of the paper is to find the SMT configuration, based on morpho-syntactic information, with the best translation results, when translating from English to the highly inflectional Slovenian language. To address this issue, we added the morphological information in terms of morpho-syntactic description (MSD) tags that were attached to words. A MSD tag includes all morpho-syntactic information in position-dependent attributes. Tags were attached to words by TreeTagger. Several experiments were performed using MSD tags to improve the translation results. First, factored translation was studied, and different configurations were tested. They show that factored translation improves modeling of short distance collocations. To capture long-distance dependencies in languages, operation sequence models (OSM) were added in the second set of experiments. An additional improvement was obtained. The overall results show that the morpho-syntactic information of inflectional language is an important factor in translation. Factored translation with OSM models brought 9% relative improvement. The most successful configuration was tSaMaL-SaMaL (OSM: 0-0, 1-1, 2-2). The conclusions of our work can be applied to other Slavic languages, as they to some extent share the same morphological characteristics.