public class BestPairwiseAlignmentKernel extends KernelOnPairs
where \(softmax(a,b) = \frac{log(e^{Fa} + e^{Fb})}{F}\) with F being a large number (F=100 is enough)
where BK is another kernel the kernel on pairs relies on.
The pairwise product kernel is described in:
[Filice et al., 2015] Simone Filice, Giovanni Da San Martino and Alessandro Moschitti.
Structural Representations for Learning Relations between Pairs of Texts. In Proc. of ACL 2015.
baseKernel| Constructor and Description |
|---|
BestPairwiseAlignmentKernel() |
BestPairwiseAlignmentKernel(Kernel baseKernel,
boolean intraPairSimProduct)
Defines a Kernel operating on pairs that applies the following formula:
\(K( < x_1, x_2 >, < y_1,y_2 > ) = softmax(BK(x_1, y_1) \cdot BK(x_2, y_2) , BK(x_1, y_2) \cdot BK(x_2, y_1))\) |
| Modifier and Type | Method and Description |
|---|---|
boolean |
getIntraPairSimProduct() |
float |
kernelComputationOverPairs(Example exA1,
Example exA2,
Example exB1,
Example exB2)
Returns the kernel computation
|
void |
setIntraPairSimProduct(boolean intraPairSimProduct)
Sets whether adding or not to the kernel combination an extra term equivalent to the
multiplication of the intra-pair similarities, i.e.:
\(BK(x_1,x_2) \cdot BK(y_1,y_2)\)
|
kernelComputationgetBaseKernel, setBaseKerneldisableCache, getKernelCache, getKernelComputations, getNumberOfHits, getNumberOfMisses, getSquaredNormCache, innerProduct, load, reset, save, setKernelCache, setSquaredNormCache, squaredNorm, squaredNormOfTheDifferencepublic BestPairwiseAlignmentKernel(Kernel baseKernel, boolean intraPairSimProduct)
where \(softmax(a,b) = \frac{log(e^{Fa} + e^{Fb})}{F}\) with F being a large number (F=100 is enough)
baseKernel - the base kernel BKintraPairSimProduct - whether adding or not the following term to K: \(BK(x_1,x_2) \cdot BK(y_1,y_2)\)
NOTE: the additional intra-pair similarity term corresponds to adding a feature \(BK(x_1,x_2)\) to the example \(< x_1, x_2 >\). Instead of enabling this additional term, it would be more efficient to explicitly adding a new vector representation to example that includes that feature. This can be easily done using the manipulate method of PairSimilarityExtractor (it is included in vector-representation)
public BestPairwiseAlignmentKernel()
public boolean getIntraPairSimProduct()
public void setIntraPairSimProduct(boolean intraPairSimProduct)
NOTE: the additional intra-pair similarity term corresponds to adding a feature \(BK(x_1,x_2)\) to the example \(< x_1, x_2 >\). Instead of enabling this additional term, it would be more efficient to explicitly adding a new vector representation to example that includes that feature. This can be easily done using the manipulate method of PairSimilarityExtractor (it is included in vector-representation)
intraPairSimProduct - whether adding or not the
multiplication of the intra-pair similaritiespublic float kernelComputationOverPairs(Example exA1, Example exA2, Example exB1, Example exB2)
KernelOnPairskernelComputationOverPairs in class KernelOnPairsexA1 - the first element of the first pairexA2 - the second element of the first pairexB1 - the first element of the second pairexB2 - the second element of the second pairCopyright © 2015 Semantic Analytics Group @ Uniroma2. All rights reserved.