Lattice Desegmentation for Statistical Machine Translation

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 100–110, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Lattice Desegmentation for Statistical Machine Translation Mohammad Salameh† Colin Cherry‡ Grzegorz Kondrak† †Department of Computing Science ‡National Research Council Canada University of Alberta 1200 Montreal Road Edmonton, AB, T6G 2E8, Canada Ottawa, ON, K1A 0R6, Canada {msalameh,gkondrak}@ualberta.ca Colin.Cherry@nrc-cnrc.gc.ca Abstract Morphological segmentation is an effec- tive sparsity reduction strategy for statis- tical machine translation (SMT) involv- ing morphologically complex languages. When translating into a segmented lan- guage, an extra step is required to deseg- ment the output; previous studies have de- segmented the 1-best output from the de- coder. In this paper, we expand our trans- lation options by desegmenting n-best lists or lattices. Our novel lattice desegmenta- tion algorithm effectively combines both segmented and desegmented views of the target language for a large subspace of possible translation outputs, which allows for inclusion of features related to the de- segmentation process, as well as an un- segmented language model (LM). We in- vestigate this technique in the context of English-to-Arabic and English-to-Finnish translation, showing significant improve- ments in translation quality over deseg- mentation of 1-best decoder outputs. 1 Introduction Morphological segmentation is considered to be indispensable when translating between English and morphologically complex languages such as Arabic. Morphological complexity leads to much higher type to token ratios than English, which can create sparsity problems during translation model estimation. Morphological segmentation addresses this issue by splitting surface forms into meaningful morphemes, while also performing or- thographic transformations to further reduce spar- sity. For example, the Arabic noun ����� lldwl “to the countries” is segmented as l+ “to” Aldwl “the countries”. When translating from Arabic, this segmentation process is performed as input preprocessing and is otherwise transparent to the translation system. However, when translating into Arabic, the decoder produces segmented out- put, which must be desegmented to produce read- able text. For example, l+ Aldwl must be con- verted to lldwl. Desegmentation is typically performed as a post-processing step that is independent from the decoding process. While this division of labor is useful, the pipeline approach may prevent the de- segmenter from recovering from errors made by the decoder. Despite the efforts of the decoder’s various component models, the system may pro- duce mismatching segments, such as s+ hzymp, which pairs the future particle s+ “will” with a noun hzymp “defeat”, instead of a verb. In this sce- nario, there is no right desegmentation; the post- processor has been dealt a losing hand. In this work, we show that it is possible to maintain the sparsity-reducing benefit of segmen- tation while translating directly into unsegmented text. We desegment a large set of possible de- coder outputs by processing n-best lists or lat- tices, which allows us to consider both the seg- mented and desegmented output before locking in the decoder’s decision. We demonstrate that sig- nificant improvements in translation quality can be achieved by training a linear model …

Content truncated to 500 words