Citation Context

... in a separate unconstraint track submission — the use of a huge language model trained on 126 billion tokens with a new training tool (Section 4). 1 Initial System Development We start with systems (=-=Haddow and Koehn, 2012-=-) that we developed for the 2012 Workshop on Statistical Machine Translation (Callison-Burch et al., 2012). The notable features of these systems are: • Moses phrase-based models with mostly default s...

Citation Context

...nts two problems. First, it has a single reference, sometimes of lower quality than the multiple references in tuning sets from MT competitions. Second, large bitexts often comprise many text genres (=-=Haddow and Koehn, 2012-=-), a virtue for classical dense MT models but a curse for high dimensional models: bitext tuning can lead to a significant domain adaptation problem when evaluating on standard test sets. Our analysis...

Citation Context

...dev and test sets, resulting in a domain mismatch. 203Therefore, we are actually comparing a smaller indomain tuning set with a larger out-of-domain set. While this domain adaptation is problematic (=-=Haddow and Koehn, 2012-=-), the ability to discriminatively tune on larger sets remains highly desirable. In terms of running time, we observe that the algorithm scales linearly with respect to input size, regardless of the f...

Citation Context

...ering setups, selecting 10%, 20% and 50% of the parallel out-of-domain data. We also used the filtered target sides of the parallel data for building language models. Another approach is described in =-=[15]-=- (labelled x+yE there and in+outE here) and modifies the IN+OUT phrase tables by replacing all scores of phrase pairs found in the in-domain data by the values estimated on in-domain data only. The id...

Citation Context

... made available through the Opus project [10]. Combination of corpora to improve the translation model has been explored with relative success in the past. For the NewsCommentary and OpenSrt corpora, =-=[11]-=- explore different ways to mix the phrase-table to adapt the Europarl corpus. For the Arabic-English IWSLT data, [12] achieve a relative improvement of 0.7 BLEU by mixing phrases from UN and IWSLT dat...

Citation Context

...irst setting, the goal is to effectively make use of both the old-domain and the new-domain parallel training corpora (Civera and Juan, 2007; Foster and Kuhn, 2007; Foster et al., 2010; Haddow, 2013; =-=Haddow and Koehn, 2012-=-; Koehn and Schroeder, 2007). In the second setting, it has been shown that, in some cases, training a translation model on a subset of new-domain parallel training data within a larger training corpu...

Citation Context

...ual target language corpus for use as language model training data (Lin et al., 1997; Klakow, 2000; Gao et al., 2002; Moore and Lewis, 2010; Mansour et al., 2011). Translation modeling research has typically assumed that either (1) two parallel datasets are available, one in the old domain and one in the new, or (2) a large, mixed-domain parallel training corpus is available. In the first setting, the goal is to effectively make use of both the old-domain and the new-domain parallel training corpora (Civera and Juan, 2007; Koehn and Schroeder, 2007; Foster and Kuhn, 2007; Foster et al., 2010; Haddow and Koehn, 2012; Haddow, 2013). In the second setting, it has been shown that, in some cases, training a translation model on a subset of newdomain parallel training data within a larger training corpus can be more effective than using the complete dataset (Mansour et al., 2011; Axelrod et al., 2011; Sennrich, 2012; Gasco et al., 2012). For many language pairs and domains, no newdomain parallel training data is available. Wu et al. (2008) machine translate new-domain source language monolingual corpora and use the synthetic parallel corpus as additional training data. Daume and Jagarlamudi (2011), Zhang an...