ISI’s 2005 Statistical Machine Translation Entries

ISI’s 2005 Statistical Machine Translation Entries Steve DeNeefe, Kevin Knight Information Sciences Institute and Department of Computer Science The Viterbi School of Engineering, University of Southern California 4676 Admiralty Way, Suite 1001 Marina del Rey, CA 90292 sdeneefe@isi.edu, knight@isi.edu Abstract ISI entered two statistical machine translation systems in the IWSLT evaluation this year: one was phrase-based and the other syntax-based. The syntax-based system represents the results of a current research effort, while the phrase-based system is representative of the current techniques in state-of- the-art machine translation. This paper primarily describes the syntax-based system and its comparison to the phrase- based system. We will give a brief overview of the com- ponents of the systems and discuss the performance on the IWSLT development data, the evaluation results, and some post-evaluation results. 1. Introduction Statistical phrase-based machine translation is currently the state-of-the-art in many translation tasks, achieving results often surpassing other methods [1]. Systems that strive to im- prove on the current results of phrase-based machine trans- lation often incorporate a higher-level notion of the struc- ture in language, sometimes just the hierarchical structure [2] and other times a fully syntactic model. For the past several years, the ISI/USC machine translation group has been inves- tigating how to use syntactic information to improve trans- lation quality beyond the capability of our existing phrase- based translation system and in the process created a new syntax-based translation system. Both systems have similar- ities: they are both statistical and trained on bilingual parallel data, both combine their translation model with several other knowledge sources in a log-linear manner, and both require parameter tuning to determine the weights of the individual components. The syntax-based system is different in two main respects: the translation model incorporates syntactic structure on the target language side (in our case, English), and the decoder uses a parser-like method to create syntactic trees as output hypotheses. 1.1. Syntax-based Translation Model Simply put, our syntax model translates phrases in the source language into syntactic chunks in the target language. For example, when translating from Chinese into English, our system learns simple rules that translate words or phrases, such as NPB(PRP(I)) ↔ · · NN(hotel) ↔ Ë· Ë Ë¡ ¡ ¡ NP-C(NPB(DT(this) NN(address))) ↔ Y Y YÇ Ç Ç • • •Œ Œ Œ It also learns phrases with “holes” in the source language (represented here by the variable x0), as long as they conform to a syntactic structure in the target language NP-C(NPB(PRP$(my) x0:NN)) ↔ · · · { { NP-C(NPB(PRP$(my) x0:NN)) ↔ ·{ x0 · · x0 PP(TO(to) NP-C(NPB(x0:NNP NNP(park)))) ↔ V V V x0 Ú Ú ÚÉ É É Other rules bring together already translated phrases, such as the following rules which take a translated verb next to a translated noun phrase and combine them together into a verb phrase: VP(x0:VBZ x1:NP-C) ↔ x0 x1 VP(x0:VBZ x1:NP-C) ↔ x1 x0 The first rule combines the pair in order. The second takes a noun phrase located before a verb, switches the order, …

Content truncated to 500 words