
epic.parser.projections.EnumeratedAnchoring.scala Maven / Gradle / Ivy
package epic.parser
package projections
/*
Copyright 2012 David Hall
Licensed under the Apache License, Version 2.0 (the "License")
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/
import java.io._
import projections.AnchoredForestProjector.ForestData
import breeze.collection.mutable.{OpenAddressHashArray, TriangularArray}
import epic.lexicon.Lexicon
import epic.constraints.ChartConstraints
import epic.trees.Span
import java.util
/**
* Creates a locally-normalized anchored PCFG from some refined forest.
* @author dlwh
*/
case class AnchoredPCFGProjector[L, W](threshold: Double = Double.NegativeInfinity) extends ChartProjector[L, W] {
type MyAnchoring = EnumeratedAnchoring[L, W]
private def normalize(grammar: RuleTopology[L], ruleScores: OpenAddressHashArray[Double], totals: OpenAddressHashArray[Double]):OpenAddressHashArray[Double] = {
if(ruleScores eq null) null
else {
val r = new OpenAddressHashArray[Double](ruleScores.length, Double.NegativeInfinity, ruleScores.activeSize)
for( (rule, score) <- ruleScores.activeIterator) {
val parent = grammar.parent(rule)
if(score > 0)
r(rule) = math.log(score) - math.log(totals(parent))
}
r
}
}
private def logify(ruleScores: OpenAddressHashArray[Double]):OpenAddressHashArray[Double] = {
if(ruleScores eq null) null
else {
val r = new OpenAddressHashArray[Double](ruleScores.length, Double.NegativeInfinity, ruleScores.activeSize)
for( (rule, score) <- ruleScores.activeIterator) {
r(rule) = math.log(score)
}
r
}
}
protected def createAnchoring(charts: ParseMarginal[L, W], ruleData: ForestData, sentProb: Double) = {
val AnchoredForestProjector.ForestData(lexicalScores, unaryScores, totalsUnaries, binaryScores, totalsBinaries) = ruleData
val normUnaries:Array[OpenAddressHashArray[Double]] = for((ruleScores, totals) <- unaryScores zip totalsUnaries) yield {
normalize(charts.topology, ruleScores, totals)
}
val normBinaries:Array[Array[OpenAddressHashArray[Double]]] = for ((splits, totals) <- binaryScores zip totalsBinaries) yield {
if(splits eq null) null
else for(ruleScores <- splits) yield normalize(charts.topology, ruleScores, totals)
}
val sparsity = charts.anchoring.sparsityPattern
new EnumeratedAnchoring(charts.topology, charts.lexicon, charts.words, lexicalScores.map(logify), normUnaries, normBinaries, sparsity)
}
}
/**
* Creates anchorings for a set of trees from some parser using p(rule | sentence) marginals.
* @author dlwh
*/
@SerialVersionUID(469174684243960202L)
case class AnchoredRuleMarginalProjector[L, W](threshold: Double = Double.NegativeInfinity) extends ChartProjector[L, W] {
private def normalize(ruleScores: OpenAddressHashArray[Double]):OpenAddressHashArray[Double] = {
if(ruleScores eq null) null
else {
val r = new OpenAddressHashArray[Double](ruleScores.length, Double.NegativeInfinity, ruleScores.activeSize)
for( (rule, score) <- ruleScores.activeIterator) {
r(rule) = math.log(score)
}
r
}
}
type MyAnchoring = EnumeratedAnchoring[L, W]
protected def createAnchoring(charts: ParseMarginal[L, W],
ruleData: ForestData,
sentProb: Double) = {
val AnchoredForestProjector.ForestData(lexicalScores, unaryScores, _, binaryScores, _) = ruleData
val normUnaries:Array[OpenAddressHashArray[Double]] = unaryScores.map(normalize)
val normBinaries:Array[Array[OpenAddressHashArray[Double]]] = for (splits <- binaryScores) yield {
if(splits eq null) null
else splits.map(normalize)
}
val sparsity = charts.anchoring.sparsityPattern
new EnumeratedAnchoring(charts.topology, charts.lexicon, charts.words, lexicalScores.map(normalize), normUnaries, normBinaries, sparsity)
}
}
/**
* An [[epic.parser.projections.EnumeratedAnchoring]] is a compressed forest, with potentially distinct
* rule and span scores for every possible anchoring into the sentence. Commonly produced by either
* an [[epic.parser.projections.AnchoredPCFGProjector]] (for PCFG forests) or an [[epic.parser.projections.AnchoredRuleMarginalProjector]]
* (for rule marginal forests).
* @tparam L
* @tparam W
*/
@SerialVersionUID(3)
case class EnumeratedAnchoring[L, W](topology: RuleTopology[L],
lexicon: Lexicon[L, W],
words: IndexedSeq[W],
spanScores: Array[OpenAddressHashArray[Double]], // triangular index -> label -> score
// (begin, end) -> rule -> score
unaryScores: Array[OpenAddressHashArray[Double]],
// (begin, end) -> (split-begin) -> rule -> score
binaryScores: Array[Array[OpenAddressHashArray[Double]]],
sparsityPattern: ChartConstraints[L]) extends UnrefinedGrammarAnchoring[L, W] with Serializable {
override def addConstraints(cs: ChartConstraints[L]): UnrefinedGrammarAnchoring[L, W] = copy(sparsityPattern = sparsityPattern & cs)
/**
* Computes the pointwise division of two grammars, augmenting
* their refinement space to reflect this. If they share the same annotationTag,
* (assuming it's non-negative) they will share their state space. (That is,
* they will have the same annotations.)
* @param other
* @return
*/
override def /(other: GrammarAnchoring[L, W]): GrammarAnchoring[L, W] = {
other match {
case that: EnumeratedAnchoring[L, W] => EnumeratedAnchoring.divide(this, that)
case _ => super./(other)
}
}
/**
* Computes the point-wise division of this grammar with some other grammar.
*
* Note that scores are in log space, so we actually subtract scores.
* @param other
* @return
*/
override def /(other: UnrefinedGrammarAnchoring[L, W]): UnrefinedGrammarAnchoring[L, W] = {
other match {
case that: EnumeratedAnchoring[L, W] => EnumeratedAnchoring.divide(this, that)
case _ => super./(other)
}
}
def scoreUnaryRule(begin: Int, end: Int, rule: Int) = {
val forSpan = unaryScores(TriangularArray.index(begin, end))
if(forSpan eq null) Double.NegativeInfinity
else forSpan(rule)
}
def scoreBinaryRule(begin: Int, split: Int, end: Int, rule: Int) = {
val ti = TriangularArray.index(begin, end)
val forSpan = binaryScores(ti)
val cached = checkCache(split, rule, ti)
if(!java.lang.Double.isNaN(cached)) {
cached
} else if(forSpan eq null) {
Double.NegativeInfinity
} else {
val forSplit = forSpan(split - begin)
val result = if(forSplit eq null) Double.NegativeInfinity
else forSplit(rule)
updateCache(split, rule, ti, result)
result
}
}
def scoreSpan(begin: Int, end: Int, tag: Int): Double = {
val scores = spanScores(TriangularArray.index(begin, end))
if(scores ne null) scores(tag)
else Double.NegativeInfinity
}
// (1 entry for each position (a split point), an entry has a rule index, a begin/end pair, and a score
private val cache = new Array[Int](length * (1 + 1 + 2))
util.Arrays.fill(cache, -1) //
private def checkCache(splitPoint: Int, rule: Int, ti: Int) = {
val crule = cache(splitPoint * 4)
val cti = cache(splitPoint * 4 + 1)
if(rule == crule && cti == ti) {
java.lang.Double.longBitsToDouble(Span(cache(splitPoint * 4 + 2), cache(splitPoint * 4 + 3)).encoded)
} else {
Double.NaN
}
}
private def updateCache(splitPoint: Int, rule: Int, ti: Int, score: Double):Unit = {
cache(splitPoint * 4) = rule
cache(splitPoint * 4 + 1) = ti
val asSpan = new Span(java.lang.Double.doubleToRawLongBits(score))
cache(splitPoint * 4 + 2) = asSpan.begin
cache(splitPoint * 4 + 3) = asSpan.end
}
}
object EnumeratedAnchoring {
def divide[L, W](a: EnumeratedAnchoring[L, W], b: EnumeratedAnchoring[L, W]): EnumeratedAnchoring[L, W] = {
val newSpanScores = Array.tabulate(a.spanScores.length) { i =>
val oldA = a.spanScores(i)
val oldB = b.spanScores(i)
if(null == oldA || null == oldB) {
null
} else {
doDivide(oldA, oldB)
}
}
val newUnaryScores = Array.tabulate(a.unaryScores.length) { i =>
val oldA = a.unaryScores(i)
val oldB = b.unaryScores(i)
if(null == oldA || null == oldB) {
null
} else {
doDivide(oldA, oldB)
}
}
val newBinaryScores = Array.tabulate(a.binaryScores.length) { i =>
val aArray = a.binaryScores(i)
val bArray = b.binaryScores(i)
if(null == aArray || null == bArray) {
null
} else {
Array.tabulate(aArray.length) { split =>
val oldA = aArray(split)
val oldB = bArray(split)
if(null == oldA || null == oldB) {
null
} else {
doDivide(oldA, oldB)
}
}
}
}
a.copy(spanScores = newSpanScores, unaryScores=newUnaryScores, binaryScores = newBinaryScores)
}
private def doDivide(a: OpenAddressHashArray[Double], b: OpenAddressHashArray[Double]) = {
if(a == null || b == null) {
null
} else {
val oah = new OpenAddressHashArray[Double](a.size, a.default, a.activeSize min b.activeSize)
var off = 0
while(off < a.iterableSize) {
if(a.isActive(off)) {
val aa = a.valueAt(off)
val ii = a.indexAt(off)
val bb = b(ii)
if(aa != Double.NegativeInfinity && bb != Double.NegativeInfinity) {
oah(ii) = aa - bb
}
}
off += 1
}
oah
}
}
}
© 2015 - 2025 Weber Informatics LLC | Privacy Policy