Feature-Rich Phrase-based Translation: Stanford University’s Submission to the WMT 2013 Translation Task

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 148–153, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics Feature-Rich Phrase-based Translation: Stanford University’s Submission to the WMT 2013 Translation Task Spence Green, Daniel Cer, Kevin Reschke, Rob Voigt*, John Bauer Sida Wang, Natalia Silveira†, Julia Neidert and Christopher D. Manning Computer Science Department, Stanford University *Center for East Asian Studies, Stanford University †Department of Linguistics, Stanford University {spenceg,cerd,kreschke,robvoigt,horatio,sidaw,natalias,jneid,manning}@stanford.edu Abstract We describe the Stanford University NLP Group submission to the 2013 Workshop on Statistical Machine Translation Shared Task. We demonstrate the effectiveness of a new adaptive, online tuning algorithm that scales to large feature and tuning sets. For both English-French and English-German, the algorithm produces feature-rich mod- els that improve over a dense baseline and compare favorably to models tuned with established methods. 1 Introduction Green et al. (2013b) describe an online, adaptive tuning algorithm for feature-rich translation mod- els. They showed considerable translation quality improvements over MERT (Och, 2003) and PRO (Hopkins and May, 2011) for two languages in a research setting. The purpose of our submission to the 2013 Workshop on Statistical Machine Trans- lation (WMT) Shared Task is to compare the algo- rithm to more established methods in an evaluation. We submitted English-French (En-Fr) and English- German (En-De) systems, each with over 100k fea- tures tuned on 10k sentences. This paper describes the systems and also includes new feature sets and practical extensions to the original algorithm. 2 Translation Model Our machine translation (MT) system is Phrasal (Cer et al., 2010), a phrase-based system based on alignment templates (Och and Ney, 2004). Like many MT systems, Phrasal models the predictive translation distribution p(e|f; w) directly as p(e|f; w) = 1 Z(f) exp � w⊤φ(e, f) � (1) where e is the target sequence, f is the source se- quence, w is the vector of model parameters, φ(·) is a feature map, and Z(f) is an appropriate nor- malizing constant. For many years the dimension of the feature map φ(·) has been limited by MERT, which does not scale past tens of features. Our submission explores real-world translation quality for high-dimensional feature maps and as- sociated weight vectors. That case requires a more scalable tuning algorithm. 2.1 Online, Adaptive Tuning Algorithm Following Hopkins and May (2011) we cast MT tun- ing as pairwise ranking. Consider a single source sentence f with associated references e1:k. Let d be a derivation in an n-best list of f that has the target e = e(d) and the feature map φ(d). Define the linear model score M(d) = w · φ(d). For any derivation d+ that is better than d− under a gold metric G, we desire pairwise agreement such that G � e(d+), e1:k� > G � e(d−), e1:k� ⇐⇒ M(d+) > M(d−) Ensuring pairwise agreement is the same as ensur- ing w · [φ(d+) − φ(d−)] > 0. For learning, we need to select derivation pairs (d+, d−) to compute difference vectors x+ = φ(d+) − φ(d−). Then …

Content truncated to 500 words