Citation Context

...ang et al., 2006; Watanabe et al., 2007), are recognized to be effective, but require substantial implementation efforts due to difficulties with parallelization. Pairwise ranking (Shen et al., 2004; =-=Hopkins and May, 2011-=-) recasts tuning as classification, and can be very easy to implement, as it fits nicely into the established MERT infrastructure. The MERT algorithm optimizes linear weights relative to a collection ...

Citation Context

...MT system relies on a training algorithm 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 learning algorithms—e.g., in the general framework of empirical risk minimization (Vapnik, 1998)—their procedural specifications have made this diff...

Citation Context

...mpression methods (Guthrie and Hepple, 2010; Talbot and Osborne, 2007) and distributed language models (Brants et al., 2007) could reduce hardware requirements. Feature weights were retuned with PRO (=-=Hopkins and May, 2011-=-) for Czech-English and batch MIRA (Cherry and Foster, 2012) for French-English and Spanish-English because these worked best for the baseline. Uncased BLEU scores on the 2013 test set are shown in Ta...

Citation Context

...it is only recently that some of them have gained popularity in the community, most notably, the margin infused relaxed algorithm (MIRA) (Chiang et al., 2008) and pairwise ranking optimization (PRO) (=-=Hopkins and May, 2011-=-). While the number of parameters that an optimizer can handle has become a major concern recently, there are many other important aspects that researchers have paid attention to, e.g., the performanc...

Citation Context

...red to the other models: each phrase table has its own feature weights, which means five additional features. It is well known that MERT cannot handle too many features (Chiang, Knight, & Wang, 2009; =-=Hopkins & May, 2011-=-), and we believe this is our case as it takes 30–35 iterations to finish, while the other methods normally only need 7–8 iterations. A closer look at MERT revealed two further issues: (1) The n-best ...

Citation Context

...erich MT systems that is effective yet comparatively easy to implement. The algorithm scales to millions of features and large tuning sets. It optimizes a logistic objective identical to that of PRO (=-=Hopkins and May, 2011-=-) with stochastic gradient descent, although other objectives are possible. The learning rate is set adaptively using AdaGrad (Duchi et al., 2011), which is particularly effective for the mixture of d...

Citation Context

...omputes the hypothesis features described above, given a set of experts E. The weight vector w is tuned on a development set of error-annotated sentences using the PRO ranking optimization algorithm (=-=Hopkins and May, 2011-=-). 1 1 We also experimented with the MERT algorithm (Och, 2003) but found that PRO achieved better results. PRO performs decoder parameter tuning through a pair-wise ranking approach. The algorithm st...

Citation Context

...egrated with both MERT implementation in Moses (Koehn (9) 91PBML 97 APRIL 2012 et al., 2007) and zMERT (Zaidan, 2009). Recently we have also added an implementation of Pairwise Ranking Optimization (=-=Hopkins and May, 2011-=-) (PRO) within Kriya for tuning the feature weights. 5. Experiments In this section we present experiments to evaluate Kriya on several language pairs. We use five different language pairs in our expe...

Citation Context

...rmore, we would like to be able to still target the BLEU measure of translation quality during learning. While large-scale discriminative training for machine translation is a widely studied problem (=-=Hopkins and May, 2011-=-; Li and Eisner, 2009; Devlin, 2009; Blunsom et al., 2008; Watanabe et al., 2007; Arun and Koehn, 2007; Liang et al., 2006), no tractable algorithm exists for learning a large number of feature weight...

Citation Context

...of discourselevel corpus annotations such as coreference links. These annotations can then be accessed by the feature models. To allow for additional targetlanguage information such as morphological features of target words, Docent can handle simple word-level annotations that are encoded in the phrase table in the same way as target language factors in Moses. In order to optimise feature weights we have adapted the Moses tuning infrastructure to Docent. In this way we can take advantage of all its features, for instance using different optimisation algorithms such as MERT (Och, 2003) or PRO (Hopkins and May, 2011), and selective tuning of a subset of features. Since document features only give meaningful scores on the document level and not on the sentence level, we naturally perform optimisation on document level, which typically means that we need more data than for the optimisation of sentence-based decoding. The results we obtain are relatively stable and competitive with sentence-level optimisation of the same models (Stymne et al., 2013a). 4 Implementing Feature Models Efficiently While translating a document, the local search decoder attempts to make a great number of moves. For each move, a sco...