Structured Ramp Loss Minimization for Machine Translation

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 221–231, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Structured Ramp Loss Minimization for Machine Translation Kevin Gimpel and Noah A. Smith Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213, USA {kgimpel,nasmith}@cs.cmu.edu Abstract This paper seeks to close the gap between training algorithms used in statistical machine translation and machine learning, specifically the framework of empirical risk minimization. We review well-known algorithms, arguing that they do not optimize the loss functions they are assumed to optimize when applied to machine translation. Instead, most have im- plicit connections to particular forms of ramp loss. We propose to minimize ramp loss di- rectly and present a training algorithm that is easy to implement and that performs compa- rably to others. Most notably, our structured ramp loss minimization algorithm, RAMPION, is less sensitive to initialization and random seeds than standard approaches. 1 Introduction Every statistical MT system relies on a training al- gorithm to fit the parameters of a scoring function to examples from parallel text. Well-known examples include MERT (Och, 2003), MIRA (Chiang et al., 2008), and PRO (Hopkins and May, 2011). While such procedures can be analyzed as machine learn- ing algorithms—e.g., in the general framework of empirical risk minimization (Vapnik, 1998)—their procedural specifications have made this difficult. From a practical perspective, such algorithms are of- ten complex, difficult to replicate, and sensitive to initialization, random seeds, and other hyperparam- eters. In this paper, we consider training algorithms that are first specified declaratively, as loss functions to be minimized. We relate well-known training algo- rithms for MT to particular loss functions. We show that a family of structured ramp loss functions (Do et al., 2008) is useful for this analysis. For example, McAllester and Keshet (2011) recently suggested that, while Chiang et al. (2008, 2009) described their algorithm as “MIRA” (Crammer et al., 2006), in fact it targets a kind of ramp loss. We note here other ex- amples: Liang et al. (2006) described their algorithm as a variant of the perceptron (Collins, 2002), which has a unique loss function, but the loss actually opti- mized is closer to a particular ramp loss (that differs from the one targeted by Chiang et al.). Och and Ney (2002) sought to optimize log loss (likelihood in a probabilistic model; Lafferty et al., 2001) but actually optimized a version of the soft ramp loss. Why isn’t the application of ML to MT more straightforward? We note two key reasons: (i) ML generally assumes that the correct output can always be scored by a model, but in MT the reference trans- lation is often unreachable, due to a model’s limited expressive power or search error, requiring the use of “surrogate” references; (ii) MT models nearly al- ways include latent derivation variables, leading to non-convex losses that have generally received little attention in ML. In this paper, we discuss how these two have caused a disconnect …

Content truncated to 500 words