Ok. It sounds like a good feature that replaces word count for confusion network/lattice input.
However, th implementation is confusing and causes the whole feature function framework to be pretty strange. I'm gonna change it so that this feature is always present for these input types. Shout if you object Hieu Sent from my flying horse On 19 Sep 2012, at 01:17 PM, Kenneth Heafield <[email protected]> wrote: > Hi, > > I don't use confusion networks, but this was a pretty important feature > for confusion network-based system combination. It would be cleaner to > just count the number of epsilons. > > Kenneth > > On 09/19/12 13:07, Hieu Hoang wrote: >> *update* >> >> It seems it allows you to have a feature that counts the number of >> 'real' words in a confusion network, ie. excluding epsilons. Does anyone >> with experience of confusion networks know if this is a good feature to >> have? There's no mention of this feature anywhere, or regression tests >> for it. >> >> if it's not particularly useful, i'm gonna delete it. >> >> On 19/09/2012 12:34, Hieu Hoang wrote: >>> Hi all >>> >>> Does anyone know the difference between 'input scores' and 'linked >>> params' for confusion networks/lattices in Moses. There's 2 variables >>> m_numInputScores >>> m_numLinkParams >>> I can't see any circumstance where they shouldn't be equal. >> >> _______________________________________________ >> Moses-support mailing list >> [email protected] >> http://mailman.mit.edu/mailman/listinfo/moses-support > _______________________________________________ > Moses-support mailing list > [email protected] > http://mailman.mit.edu/mailman/listinfo/moses-support _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
