Provides an implementation of the linear feature functions typically used in hierarchical phrase-based decoding for statistical machine translation. The following is a note from Juri describing some of the functionality of the feature functions interfaces and default abstract classes.

The equality that I intended for is ff.transitionLogP() =
ff.estimateLogP() + ff.reEstimateTransitionLogP(). The re-estimate
fixes the estimate to be the true transition cost that takes into
account the state. Before decoding the cost of applying a rule is
estimated via estimateLogP() and yields the phrasal feature costs plus
an LM estimate of the cost of the lexical portions of the rule.
transitionLogP() takes rule and state and computes everything from
scratch, whereas reEstimateTransitionLogP() adds in the cost of new
n-grams that result from combining the rule with the LM states and
subtracts out the cost of superfluous less-than-n-grams that were
overridden by the updated cost calculation.

Hope this helps.