Tuning Phrase-Based Segmented Translation for a Morphologically Complex Target Language

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 105–111, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. Tuning Phrase-Based Segmented Translation for a Morphologically Complex Target Language Stig-Arne Grönroos Department of Signal Processing and Acoustics Aalto University, Finland stig-arne.gronroos@aalto.fi Sami Virpioja Department of Computer Science Aalto University, Finland sami.virpioja@aalto.fi Mikko Kurimo Department of Signal Processing and Acoustics Aalto University, Finland mikko.kurimo@aalto.fi Abstract This article describes the Aalto Uni- versity entry to the English-to-Finnish shared translation task in WMT 2015. The system participates in the con- strained condition, but in addition we impose some further constraints, using no language-specific resources beyond those provided in the task. We use a morphological segmenter, Morfessor FlatCat, but train and tune it in an un- supervised manner. The system could thus be used for another language pair with a morphologically complex tar- get language, without needing modifi- cation or additional resources. 1 Introduction In isolating languages, such as English, suit- able smallest units of translation are easy to find using whitespace and punctuation char- acters as delimiters. This approach of us- ing words as the smallest unit of transla- tion is problematic for synthetic languages with rich inflection, derivation or compound- ing. Such languages have very large vocabu- laries, leading to sparse statistics and many out-of-vocabulary words. A synthetic language uses fewer words than an isolating language to express the same sentence, by combining several grammatical markers into each word and using compound words. This difference in granularity is prob- lematic in alignment, when a word in the iso- lating language properly aligns with only a part of a word in the synthetic language. In order to balance the number of tokens between target and source, it is often possi- ble to segment the morphologically richer side. Oversegmentation is detrimental, however, as longer windows of history need to be used, and useful phrases become more difficult to extract. It is therefore important to find a balance in the amount of segmentation. linguistically accurate segmentation may beA oversegmented for the task of translation, if some of the distinctions are either unmarked or marked in a similar way in the other lan- guage. An increase in the number of tokens means that the distance spanned by dependencies becomes longer. Recurrent Neural Network (RNN) based language models have been shown to perform well for English (Mikolov et al., 2011). Their strength lies in being the- oretically capable of modeling arbitrarily long dependencies. Moreover, a huge vocabulary is particularly detrimental for neural language models due to their computationally heavy training and need to marginalize over the whole vocabulary dur- ing prediction. As morphological segmenta- tion can reduce the vocabulary size consider- ably, using RNN language models seems even more suitable for this approach. Our system is designed for translation in the direction from a morphologically less complex to a more complex language. The opposite direction – simplifying morphology – has re- ceived more attention, especially with English as the target language. Of the target languages …

Content truncated to 500 words