Integrating Morphological Desegmentation into Phrase-based Decoding

Proceedings of NAACL-HLT 2016, pages 1175–1180, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Integrating Morphological Desegmentation into Phrase-based Decoding 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 When translating into a morphologically com- plex language, segmenting the target language can reduce data sparsity, while introducing the complication of desegmenting the system output. We present a method for decoder- integrated desegmentation, allowing features that consider the desegmented target, such as a word-level language model, to be considered throughout the entire search space. Our re- sults on a large-scale, English to Arabic trans- lation task show significant improvement over the 1-best desegmentation baseline. 1 Introduction State-of-the-art systems for translating into morpho- logically complex languages, such as Arabic, em- ploy morphological segmentation of the target lan- guage in order to reduce data sparsity and improve translation quality. For example, the Arabic word �������� ldwlthm “for their country” is segmented into the prefix +� l+ “for”, the stem ����� dwlp “coun- try” and the suffix ��+ +hm “their.” The segmen- tation sometimes involves performing orthographic normalizations, such as transforming the stem-final t to p. The result is not only a reduction in the number of word types, but also better token-to-token corre- spondence with the source language. Morphological segmentation is typically per- formed as a pre-processing step before the train- ing phase, which results in a model that translates the source language into segmented target language. Desegmentation is the process of transforming the segmented output into a readable word sequence, which can be performed using a table lookup com- bined with a small set of rules. Desegmentation is usually applied to the 1-best output of the decoder. However, this pipeline suffers from error propaga- tion: errors made during decoding cannot be cor- rected, even when desegmentation results in an ille- gal or extremely unlikely word. Two principal types of solutions have been proposed for this problem: rescoring and phrase-table desegmentation. The rescoring approach desegments either an n-best list (Oflazer and Durgar El-Kahlout, 2007) or lattice (Salameh et al., 2014), and then re-ranks with features that consider the desegmented word sequence of each hypothesis. Rescoring features include the score from an unsegmented target lan- guage model and contiguity indicators that flag target words that were translated from contiguous source tokens. Rescoring widens the desegmenta- tion pipeline, allowing desegmentation features to reduce the number of translation errors. However, these features are calculated for only a subset of the search space, and the extra rescoring step compli- cates the training and translation processes. Phrase-table desegmentation (Luong et al., 2010) also translates into a segmented target language, but alters training to perform word-boundary-aware phrase extraction. The extracted phrases are con- strained to contain only complete target words, with- out any dangling affixes. With this restriction in place, the phrase table can be desegmented before decoding begins, allowing the decoder …

Content truncated to 500 words