public class Builder<T> extends ObjectBuilds a minimal FST (maps an IntsRef term to an arbitrary output) from pre-sorted terms with outputs. The FST becomes an FSA if you use NoOutputs. The FST is written on-the-fly into a compact serialized format byte array, which can be saved to / loaded from a Directory or used directly for traversal. The FST is always finite (no cycles).
NOTE: The algorithm is described at http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.24.3698
The parameterized type T is the output type. See the subclasses of
FSTs larger than 2.1GB are now possible (as of Lucene 4.2). FSTs containing more than 2.1B nodes are also now possible, however they cannot be packed.
- WARNING: This API is experimental and might change in incompatible ways in the next release.
Constructors Constructor Description
Builder(FST.INPUT_TYPE inputType, int minSuffixCount1, int minSuffixCount2, boolean doShareSuffix, boolean doShareNonSingletonNodes, int shareMaxTailLength, Outputs<T> outputs, boolean allowFixedLengthArcs, int bytesPageBits)Instantiates an FST/FSA builder with all the possible tuning and construction tweaks.
Builder(FST.INPUT_TYPE inputType, Outputs<T> outputs)Instantiates an FST/FSA builder without any pruning.
All Methods Instance Methods Concrete Methods Modifier and Type Method Description
add(IntsRef input, T output)Add the next input/output pair.
finish()Returns final FST.
setDirectAddressingMaxOversizingFactor(float factor)Overrides the default the maximum oversizing of fixed array allowed to enable direct addressing of arcs instead of binary search.
public Builder(FST.INPUT_TYPE inputType, Outputs<T> outputs)Instantiates an FST/FSA builder without any pruning. A shortcut to
Builder(FST.INPUT_TYPE, int, int, boolean, boolean, int, Outputs, boolean, int)with pruning options turned off.
public Builder(FST.INPUT_TYPE inputType, int minSuffixCount1, int minSuffixCount2, boolean doShareSuffix, boolean doShareNonSingletonNodes, int shareMaxTailLength, Outputs<T> outputs, boolean allowFixedLengthArcs, int bytesPageBits)Instantiates an FST/FSA builder with all the possible tuning and construction tweaks. Read parameter documentation carefully.
inputType- The input type (transition labels). Can be anything from
FST.INPUT_TYPEenumeration. Shorter types will consume less memory. Strings (character sequences) are represented as
FST.INPUT_TYPE.BYTE4(full unicode codepoints).
minSuffixCount1- If pruning the input graph during construction, this threshold is used for telling if a node is kept or pruned. If transition_count(node) >= minSuffixCount1, the node is kept.
minSuffixCount2- (Note: only Mike McCandless knows what this one is really doing...)
true, the shared suffixes will be compacted into unique paths. This requires an additional RAM-intensive hash map for lookups in memory. Setting this parameter to
falsecreates a single suffix path for all input sequences. This will result in a larger FST, but requires substantially less memory and CPU during building.
doShareNonSingletonNodes- Only used if doShareSuffix is true. Set this to true to ensure FST is fully minimal, at cost of more CPU and more RAM during building.
shareMaxTailLength- Only used if doShareSuffix is true. Set this to Integer.MAX_VALUE to ensure FST is fully minimal, at cost of more CPU and more RAM during building.
outputs- The output type for each input sequence. Applies only if building an FST. For FSA, use
NoOutputs.getNoOutput()as the singleton output object.
allowFixedLengthArcs- Pass false to disable the fixed length arc optimization (binary search or direct addressing) while building the FST; this will make the resulting FST smaller but slower to traverse.
bytesPageBits- How many bits wide to make each byte block in the BytesStore; if you know the FST will be large then make this larger. For example 15 bits = 32768 byte pages.
setDirectAddressingMaxOversizingFactorOverrides the default the maximum oversizing of fixed array allowed to enable direct addressing of arcs instead of binary search.
Setting this factor to a negative value (e.g. -1) effectively disables direct addressing, only binary search nodes will be created.
- See Also:
public float getDirectAddressingMaxOversizingFactor()
public long getTermCount()
public long getNodeCount()
public long getArcCount()
public long getMappedStateCount()
public void add(IntsRef input, T output) throws IOExceptionAdd the next input/output pair. The provided input must be sorted after the previous one according to
IntsRef.compareTo(org.apache.lucene.util.IntsRef). It's also OK to add the same input twice in a row with different outputs, as long as
Outputs.merge(T, T)method. Note that input is fully consumed after this method is returned (so caller is free to reuse), but output is not. So if your outputs are changeable (eg
IntSequenceOutputs) then you cannot reuse across calls.
public FST<T> finish() throws IOExceptionReturns final FST. NOTE: this will return null if nothing is accepted by the FST.
public long fstRamBytesUsed()