Soft: Tree Kernels and multiple feature vectors in SVM-LIGHT

Thierry Hamon thierry.hamon at LIPN.UNIV-PARIS13.FR
Tue Dec 12 17:11:02 UTC 2006

Date: Mon, 11 Dec 2006 14:34:07 +0100
From: "Alessandro Moschitti" <moschitti at>
Message-ID: <00f301c71d29$0db57950$000000a0 at ParideMobile>

[apologies for cross posting]

Dear all,

I have just released the SVM-LIGHT-TK1.2 software. This allows us to
describe a classifying object using a set of trees and a set of
vectors in the input of Support Vector Machines.

Sets of trees are useful to encode different structured features,
e.g. it is possible to select different portions of a parse tree,
independently evaluate tree kernels over them and combine the obtained
contributions.  Feature Vectors are extremely important to combine
different spaces of manually designed features and are essential to
design SVM models that work on instance pairs (tuples),
e.g. re-ranking models.

The main software features are listed hereafter:

- Fast Kernel computation.

- Tree forests, i.e. a set of trees over multiple feature spaces can
  be specified in the input.

- Vector sets, i.e. multiple feature vectors over multiple feature
  spaces can be specified in the input.

- Two types of tree kernels, i.e. subset tree and subtree kernels.

- Embedded combinations of tree and vector-based kernels.

- A commented example on how to design our own kernels.

If you are interested, you can read more about the software and
download it here:

I appreciate any bug reports, requests and comments.

Best regards,


Dr. Alessandro Moschitti
Dept. of Computer Science, Systems and Production
University of Rome Tor Vergata
Via del Politecnico 1,
00133 Rome, Italy

tel  +39 06 7259 7333
fax  +39 06 72597460
e-mail: moschitti at

Message diffuse par la liste Langage Naturel <LN at>
Informations, abonnement :
English version       : 
Archives                 :

La liste LN est parrainee par l'ATALA (Association pour le Traitement
Automatique des Langues)
Information et adhesion  :

More information about the Ln mailing list