OpenGrm SFST Glossary
- backoff FST
- a canonical FST for which each state
s
that has a failure transition to a state s'
and another transition with a label x
then there is also a transition labeled with x
from s'
.
- canonical FST
- an FST for which:
- the states are sorted by input label
- there may be failure transitions but
- there is at most one such transition per state
- there are no failure-transition (and/or epsilon-transition) cycles
- no assumption is made of general determinism or what transitions must be present on failure (unlike in a canonical n-gram model).
- there may be epsilon transitions^{1} but they are treated by failure transitions as regular symbols with each instance behaving as if it is uniquely labeled (i.e, they are not constrained by failure transitions).
- faliure transition
- specially (phi) labeled transitions that are taken only when no immediate match is possible at a given state
- normalized FST
- a canonical FST for which the weights of the paths into the future from each state sum to Weight::One()^{2}
^{1}When the
phi_label
is not 0.
^{2}Computation is done using the log semiring (
Log64Weight), appropriate for negative log probabilities. The input weight type is converted to this type internally if needed (with conversion done using a
WeightConvert
functor, pre-defined for common weight types like
TropicalWeight
and
LogWeight
).