Anaphora Models and Reordering for Phrase-Based SMT

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 122–129, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics Anaphora Models and Reordering for Phrase-Based SMT Christian Hardmeier Sara Stymne J¨org Tiedemann Aaron Smith Joakim Nivre Uppsala University Department of Linguistics and Philology firstname.lastname@lingfil.uu.se Abstract We describe the Uppsala University sys- tems for WMT14. We look at the integra- tion of a model for translating pronomi- nal anaphora and a syntactic dependency projection model for English–French. Fur- thermore, we investigate post-ordering and tunable POS distortion models for English– German. 1 Introduction In this paper we describe the Uppsala University systems for WMT14. We present three different systems. Two of them are based on the document- level decoder Docent (Hardmeier et al., 2012; Hard- meier et al., 2013a). In our English–French sys- tem we extend Docent to handle pronoun anaphora, and in our English–German system we add part- of-speech phrase-distortion models to Docent. For German–English we also have a system based on Moses (Koehn et al., 2007). Again the focus is on word order, this time by using pre- and post- reordering. 2 Document-Level Decoding Traditional SMT decoders translate texts as bags of sentences, assuming independence between sen- tences. This assumption allows efficient algorithms for exploring a large search space based on dy- namic programming (Och et al., 2001). Because of the dynamic programming assumptions it is hard to directly include discourse-level and long-distance features into a traditional SMT decoder. In contrast to this very popular stack decoding approach, our decoder Docent (Hardmeier et al., 2012; Hardmeier et al., 2013a) implements a search procedure based on local search. At any stage of the search process, its search state consists of a complete document translation, making it easy for feature models to access the complete document with its current translation at any point in time. The search algorithm is a stochastic variant of standard hill climbing. At each step, it generates a successor of the current search state by randomly applying one of a set of state changing operations to a ran- dom location in the document, and accepts the new state if it has a better score than the previous state. The operations are to change the translation of a phrase, to change the word order by swapping the positions of two phrases or moving a sequence of phrases, and to resegment phrases. The initial state can either be initialized randomly, or be based on an initial run from Moses. This setup is not limited by dynamic programming constraints, and enables the use of the full translated target document to extract features. 3 English–French Our English–French system is a phrase-based SMT system with a combination of two decoders, Moses (Koehn et al., 2007) and Docent (Hardmeier et al., 2013a). The fundamental setup is loosely based on the system submitted by Cho et al. (2013) to the WMT 2013 shared task. Our phrase table is trained on data taken from the News commentary, Europarl, UN, Common crawl and 109 …

Content truncated to 500 words