The “Noisier Channel”: Translation from Morphologically Complex Languages

Proceedings of the Second Workshop on Statistical Machine Translation, pages 207–211, Prague, June 2007. c⃝2007 Association for Computational Linguistics The ‘noisier channel’: translation from morphologically complex languages Christopher J. Dyer Department of Linguistics University of Maryland College Park, MD 20742 redpony@umd.edu Abstract This paper presents a new paradigm for translation from inflectionally rich lan- guages that was used in the University of Maryland statistical machine transla- tion system for the WMT07 Shared Task. The system is based on a hierarchical phrase-based decoder that has been aug- mented to translate ambiguous input given in the form of a confusion network (CN), a weighted finite state representation of a set of strings. By treating morphologi- cally derived forms of the input sequence as possible, albeit more “costly” paths that the decoder may select, we find that sig- nificant gains (10% BLEU relative) can be attained when translating from Czech, a language with considerable inflectional complexity, into English. 1 Introduction Morphological analysis occupies a tenuous position statistical machine translation systems. Conven- tional translation models are constructed with no consideration of the relationships between lexical items and instead treat different inflected (observed) forms of identical underlying lemmas as completely independent of one another. While the variously inflected forms of one lemma may express differ- ences in meaning that are crucial to correct transla- tion, the strict independence assumptions normally made exacerbate data sparseness and lead to poorly estimated models and suboptimal translations. A va- riety of solutions have been proposed: Niessen and Ney (2001) use of morphological information to im- prove word reordering before training and after de- coding. Goldwater and McClosky (2005) show im- provements in a Czech to English word-based trans- lation system when inflectional endings are simpli- fied or removed entirely. Their method can, how- ever, actually harm performance since the discarded morphemes carry some information that may have bearing on the translation (cf. Section 3.3). To avoid this pitfall, Talbot and Osborne (2006) use a data- driven approach to cluster source-language morpho- logical variants that are meaningless in the target language, and Yang and Kirchhoff (2006) propose the use of a backoff model that uses morphologically reduced forms only when the translation of the sur- face form is unavailable. All of these approaches have in common that the decisions about whether to use morphological information are made in either a pre- or post-processing step. Recent work in spoken language translation sug- gests that allowing decisions about the use of mor- phological information to be made along side other translation decisions (i.e., inside the decoder), will yield better results. At least as early as Ney (1999), it has been shown that when translating the out- put from automatic speech regonition (ASR) sys- tems, the quality can be improved by considering multiple (rather than only a single best) transcrip- tion hypothesis. Although state-of-the-art statistical machine translation systems have conventionally as- sumed unambiguous input; recent work has demon- strated the possibility of efficient decoding of am- 207 biguous input (represented as confusion networks …

Content truncated to 500 words