The following operations are provided for SFSTs. Care must be taken that the input FSTs meet the specified requirements (e.g.
).
The binary commands typically check their input requirements are satisfied or raise an error but the C++ versions may not check for efficiency (see the source code documentation for specific cases).
Operation |
Usage |
Description |
Complexity |
Approx |
Approx(ifst, &backoff_fst, phi_label, delta) |
approximates a normalized stochastic FST wrt a provided backoff-complete FST |
same as ShortestDistance on the intersection of the input and output FSTs |
RandGen |
fst::RandGen(ifst, &ofst, fst::RandGenOptions<SFstArcSelector>(...)) |
randomly generates paths in a stochastic FST (correctly dealing with failure transitions) |
see RandGen |
GlobalNormalize |
GlobalNormalize(&fst, phi_label, delta) |
globally normalizes, when possible1, a canonical weighted FST preserving total path weights (up to a global constant) |
same as ShortestDistance |
IsCanonical |
IsCanonical(fst, phi_label) |
checks the second property here holds for a weighted FST |
Time, Space: O(V + E) |
IsNormalized |
IsNormalized(fst, phi_label, delta) |
checks the two properties here hold for a weighted FST |
Time, Space: O(V + E) |
LocalNormalize |
LocalNormalize(&fst) |
locally normalizes, when possible, a canonical weighted FST preserving each state's out-going arc weights up to a state-specific constant |
Time, Space: O(V + E) |
Count |
|
|
NGramApprox |
NGramApprox(ifst, &ofst, order, phi_label, delta) |
approximates a normalized stochastic FST as an n-gram model (having phi_labels in OpenGrm NGram format) |
same as ShortestDistance on the intersection of the input and output FSTs |
Perplexity |
Perplexity(fst, phi_label, unknown_label, unknown_class_size) |
computes perplexity for a stochastic FST |
|
PhiNormalize |
PhiNormalize(&fst, phi_label) |
normalizes, when possible, a canonical weighted FST by only modifying the failure transitions |
Time, Space: O(V + E) |
|
sfstapprox[--phi_label=$l][--delta=$d] in.fst backoff.fst out.fst |
|
|
|
sfstcount |
|
|
|
sfstngramapprox [--order=$o][--phi_label=$l][--delta=$d] in.fst out.fst |
|
|
|
sfstnormalize [--method=global] [--phi_label=$l][--delta=$d] in.fst out.fst |
|
|
|
sfstnormalize --method=phi [-phi_label=$l][--delta=$d] in.fst out.fst |
|
|
|
sfstnormalize -method=local in.fst out.fst |
|
|
|
sfstperplexity [--phi_label=$l] [-unknown_label=$u][--unknown_class_size=$s] in.fst test.far |
(test sentences are in FST archive format) |
|
|
sfstrandgen [--phi_label=$l] [--max_length=$l] [--npath=$n] [--seed=$s] in.fst out.fst |
|
|
|
sfsttrim -phi_label=$l in.fst out.fst |
|
|
Trim |
Trim(&fst, phi_label) |
removes useless states and transitions in stochastic automata (irrespective of weights) |
Time, Space: O(V + E * max-phi-order) |
Possible when the sum of weight of all successful paths from the initial state is finite (and the input is
).