Hi list,

I have added two phrase table types to Moses, internally called
MultiModel and MultiModelCounts.

These table types construct a virtual phrase table online from a vector
of component models. MultiModel so far only supports a linear
interpolation of the probabilities in the component models, but the
addition of new combination algorithms is possible. In principle, the
combination is equivalent to the linear interpolation performed in the
tmcombine scripts (already included in Moses), but because the
combination/weighting is performed during decoding, a single phrase
table can be adapted to multiple domains by simply providing different
weights for each domain. Weights can be set for each sentence through an
API hook to mosesserver.

MultiModelCounts follows the same principle of constructing a virtual
phrase table online, but does so on the basis of sufficient statistics,
phrase (pair) frequencies. While an unweighted form corresponds to a
model trained on the concatenation of all training data (excepting word
alignment, pruning and rounding differences), the frequencies of each
component models can be weighted online.

Possible uses of the two phrase table types include quick domain
adaptation in a multi-domain environment, and the quick addition/removal
of models (without a need to retrain or run offline scripts).
Applications in interactive machine translation are also conceivable
(e.g. using a component model vector that is a mix of static background
models and incrementally updateable ones).

The architecture (and a possible application) are described in an
upcoming ACL paper: Sennrich, Rico; Schwenk, Holger; Aransa, Walid
(2013). A Multi-Domain Translation Model Framework for Statistical
Machine Translation. Proceedings of ACL 2013.

A documentation is available in
http://www.statmt.org/moses/?n=Moses.AdvancedFeatures#ntoc50

Rico

_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to