Citation Context

...ears. In particular, we plan to extend the rule extraction to the produce syntax-augmented grammars [8], which have been shown to improve on both phrase-based and hierarchical models in some settings =-=[21]-=-. We also plan to implement optimizations for decoding with syntactic grammars, such as tree binarization [22]. 8. Acknowledgements This work was supported by the EuroMatrixPlus project funded by the ...

Citation Context

...and movement based on hierarchical rules extracted from aligned parallel text. The approach has been widely adopted and reported to be competitive with other large-scale data driven approaches, e.g. (=-=Zollmann et al., 2008-=-). Large-scale hierarchical SMT involves automatic rule extraction from aligned parallel text, model parameter estimation, and the use of cube pruning k-best list generation in hierarchical translatio...

Citation Context

...lity, evidence suggests there is a limit as to what can be accomplished using only simple phrases, for example, satisfactory capturing of context-sensitive reordering phenomena between language pairs =-=[2]-=-. This assertion has been acknowledged within the field as illustrated by the recent shift in focus towards more linguistically motivated models. Aside from the development of fully syntax-based model...

Citation Context

...rocessing also does not allow the interaction of long-distance reordering decisions with nearby translation decisions via the language model. In syntax-based SMT, the size of reordering is given by the span of the grammar rules. In approaches which do not use linguistic syntactic labels (such as ITG (Wu, 1997) or Hiero, where only the start symbol S and the non-terminal X are used), the maximal span size allowed in implementations is often between 10 and 15 tokens, because using wider spans has (in experiments done in the past) resulted in decreased translation quality (e.g., (Chiang, 2007)). Zollmann et al. (2008) expand the span size to 15 only for the translation of short sentences. We present work within the hierarchical phrase-based MT framework that considers rules allowed to span up to 50 words. Approaches using linguistic syntactic labels (obtained from a source language or target language parser, or both) sometimes also use such span restrictions. However, systems which use source-side-syntactic parses of the test set sometimes do not use such a restriction because they force a match with a syntactic constituent (in the source language parse). There have been many approaches looking at backing ...

Citation Context

... cannot easily compute certain important translations [48, 1, 49, 9] and that the structural information provided by modern and reliable parsers [11, 28, 27, 6] actually helps the translation process =-=[53]-=-. This development created renewed interest in tree automata [5, 8] and tree transducers [32, 14], which are finite-state devices that compute tree languages and tree transformations, respectively. Ho...

Citation Context

... 0): minimum occurrence frequency thresholds for non-lexical and lexical rules respectively. Increasing these thresholds reduces the size of the grammar, but often at the cost of translation quality (=-=Zollmann et al., 2008-=-). • -min_freq_given_src_arg (default: 0): minimum relative frequency of a rule given its labeled source. The labeling and extraction procedures defined above identify rules from the input wordaligned...

Citation Context

... be. While relative improvements over phrase-based baselines have been reported for some language pairs, those baselines seem to remain the best option for other language pairs (DeNeefe et al., 2007; =-=Zollmann et al., 2008-=-). The performance of syntax-based models is affected by errors introduced by existing imperfect syntactic parsers (Quirk and Corston-Oliver., 2006). Moreover, some non-syntactic phrases (e.g. I’m) id...

Citation Context

...grating syntactic components. This approach enables to achieve better word and component reordering than basic SMT systems, especially for structurally distant language pairs such as Chinese-English (=-=Zollmann et al., 2008-=-). Another recent trend is the addition of statistical components to RBMT systems, sometimes called hybridisation (Boitet et al. (2009); Gulati (2011)). Research systems While the way commercial trans...