Probes in a Taxonomy of Factored Phrase-Based Models

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 253–260, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Probes in a Taxonomy of Factored Phrase-Based Models ∗ Ondˇrej Bojar, Bushra Jawaid, Amir Kamran Charles University in Prague, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostransk´e n´am. 25, Praha 1, CZ-118 00, Czech Republic {bojar,jawaid,kamran}@ufal.mff.cuni.cz Abstract We introduce a taxonomy of factored phrase- based translation scenarios and conduct a range of experiments in this taxonomy. We point out several common pitfalls when de- signing factored setups. The paper also de- scribes our WMT12 submissions CU-BOJAR and CU-POOR-COMB. 1 Introduction Koehn and Hoang (2007) introduced “factors” to phrase-based MT to explicitly capture arbitrary fea- tures in the phrase-based model. In essence, input and output tokens are no longer atomic units but rather vectors of atomic values encoding e.g. the lex- ical and morphological information separately. Fac- tored translation has been successfully applied to many language pairs and with diverse types of infor- mation encoded in the additional factors, i.a. (Bojar, 2007; Avramidis and Koehn, 2008; Stymne, 2008; Badr et al., 2008; Ramanathan et al., 2009; Koehn et al., 2010; Yeniterzi and Oflazer, 2010). On the other hand, it happens quite frequently, that the factored setup causes a loss compared to the phrase-based baseline. The underlying reason is the complexity of the search space which gets boosted when the model explicitly includes detailed information, see e.g. Bo- jar and Kos (2010) or Toutanova et al. (2008). ∗ This work was supported by the project EuroMatrixPlus (FP7-ICT-2007-3-231720 of the EU and 7E09003+7E11051 of the Czech Republic) and the Czech Science Foundation grants P406/11/1499 and P406/10/P259. We are grateful for review- ers’ comments but we have to obey the 6 page limit. Thanks also to Aleˇs Tamchyna for supplementary material on MERT. Number of Number of Translation Independent Structure Steps Searches of Searches Nickname One One – Direct Several One – Single-Step Several Serial Two-Step Complex Complex Figure 1: A taxonomy of factored phrase-based models. In this paper, we first provide a taxonomy of (phrase-based) translation setups and then we exam- ine a range of sample configurations in this taxon- omy. We don’t state universal rules, because the ap- plicability of each of the setups depends very much on the particular language pair, text domain and amount of data available, but we hope to draw at- tention to relevant design decisions. The paper also serves as the description of our WMT12 submissions CU-BOJAR and CU-POOR- COMB between English and Czech. 2 A Taxonomy of Factored P-B Models Figure 1 suggests a taxonomy of various Moses se- tups. Following the definitions of Koehn and Hoang (2007), a search consists of several translation and generation steps: translation steps map source fac- tors to target factors and generation steps produce target factors from other target factors. The taxonomy is vaguely linked to the types of problems that can be expected with a given config- uration. Direct translation is likely to suffer …

Content truncated to 500 words