make

How to Use

SVM^cfg consists of a learning module (svm_cfg_learn) and a classification module (svm_cfg_classify). The classification module can be used to apply the learned model to new examples. See also the examples below for how to use svm_cfg_learn and svm_cfg_classify.

Usage is much like SVM^light. You call it like

      svm_cfg_learn -c 1.0 example_file model_file

General Options:
         -?          -> this help
         -v [0..3]   -> verbosity level (default 1)
         -y [0..3]   -> verbosity level for svm_light (default 0)
Learning Options:
         -c float    -> C: trade-off between training error
                        and margin (default 0.01)
         -p [1,2]    -> L-norm to use for slack variables. Use 1 for L1-norm,
                        use 2 for squared slacks. (default 1)
         -o [1,2]    -> Rescaling method to use for loss.
                        1: slack rescaling
                        2: margin rescaling
                        (default 2)
         -l [0,1]    -> Loss function to use.
                        0: zero/one loss
                        1: (1-F_1) as the loss as in [1]
                        (default 0)
Optimization Options (see [1][4]):
         -w [0,..,9] -> choice of structural learning algorithm (default 4):
                        0: n-slack algorithm described in [1]
                        1: n-slack algorithm with shrinking heuristic
                        2: 1-slack algorithm (primal) described in [4]
                        3: 1-slack algorithm (dual) described in [4]
                        4: 1-slack algorithm (dual) with constraint cache [4]
                        9: custom algorithm in svm_struct_learn_custom.c
         -e float    -> epsilon: allow that tolerance for termination
                        criterion (default 0.100000)
         -k [1..]    -> number of new constraints to accumulate before
                        recomputing the QP solution (default 100) (-w 0 and 1 only)
         -f [5..]    -> number of constraints to cache for each example
                        (default 5) (used with -w 4)
         -b [1..100] -> percentage of training set for which to refresh cache
                        when no epsilon violated constraint can be constructed
                        from current cache (default 100%) (used with -w 4)
SVM-light Options for Solving QP Subproblems (see SVM-light web page):
         -n [2..q]   -> number of new variables entering the working set
                        in each svm-light iteration (default n = q).
                        Set n < q to prevent zig-zagging.
         -m [5..]    -> size of svm-light cache for kernel evaluations in MB
                        (default 40) (used only for -w 1 with kernels)
         -h [5..]    -> number of svm-light iterations a variable needs to be
                        optimal before considered for shrinking (default 100)
         -# int      -> terminate svm-light QP subproblem optimization, if no
                        progress after this number of iterations.
                        (default 100000)
Kernel Options:
         -t int      -> type of kernel function:
                        0: linear (default)
                        1: polynomial (s a*b+c)^d
                        2: radial basis function exp(-gamma ||a-b||^2)
                        3: sigmoid tanh(s a*b + c)
                        4: user defined kernel from kernel.h
         -d int      -> parameter d in polynomial kernel
         -g float    -> parameter gamma in rbf kernel
         -s float    -> parameter s in sigmoid/poly kernel
         -r float    -> parameter c in sigmoid/poly kernel
         -u string   -> parameter of user defined kernel
Output Options:
         -a string   -> write all alphas to this file after learning
                        (in the same order as in the training set)
Application-Specific Options:
         --s [1..]   -> Maximum length of sentences to use (default 10)
         --p {0,1}   -> Use parent annotation for rules (default 0)
         --a {0,1}   -> Use borders as feature (default 0)
         --b {0,1}   -> Use length of parent span as feature (default 0)
         --c {0,1}   -> Use length of childrens spans as features (default 0)
         --d {0,1}   -> Use difference between children spans as feature
                        (default 0)
         -l {0,1}    NOTE: Added loss type '1' in addition to the 0/1-loss '0'.
                     Loss type '1' uses (1-F_1) as the loss as in [1].

The option --s allows you to specify a maximum sentence length, so that all longer sentences are ignored in the experiment. Option --p turns on parent annotation of rules as described in [2]. The remaining applications-specific option (--a to --d) select different types of features to include in the representation. For more details on the meaning of the remaining options consult the description of SVM^struct , the description of SVM^light, and references [1][3][4].

The input file example_file contains the training examples. The file format is the same as in the Penn Treebank with the slight modifications described in [2] (e.g. removal of epsilon rules). For example, the following bracketed tree is an example with the terminal symbols in the leaves.

('root' (S (NP (DET the) (N cat)) (VP (V ran))))

Note that all trees have to start with the special root symbol 'root'. Note also that the current version of the code ignores the terminals, but starts with the pre-terminals.

The result of svm_cfg_learn is the model which is learned from the training data in example_file. The model is written to two files starting with model_file and having suffixes .svm and .grammar. To make predictions on test examples, svm_cfg_classify reads this file. svm_cfg_classify is called as follows:

For all test examples in test_example_file the predicted parse trees are written to output_file. There is one line per test example in output_file in the same order as in test_example_file. The test file has the same format as the training file (i.e. if you do not know the correct parse tree for the test example, you have to make up a bogus one).

Getting started: an Example Problem

You will find a small example problem with 29 examples from the Penn Treebank at

Download this file into your svm_cfg directory and unpack it with

This will create a subdirectory example5. There are 29 examples in the file train.dat. To run the example, execute the commands:

The first command trains using all sentences of length at most 40, the regularization parameter C equal to 1.0, the F₁-loss, and features a,b,c. The second command reclassifies the training set. The accuracy and the average loss on the training set is printed to stdout.

Disclaimer

This software is free only for non-commercial use. It must not be distributed without prior permission of the author. The author is not responsible for implications from the use of this software.

Known Problems

• none.

History

V3.00 - V3.10

• Included version V3.10 of SVM^struct, which makes it substantially faster.
• Fixed bug in RBF Kernel.
• Fixed precision issues on 64-bit AMD and Intel machines.
• Source code for SVM^cfg V3.00.

V2.01 - V3.00

• New training algorithm based on equivalent 1-slack reformulation of the training problem. This makes training on the linear kernel several orders of magnitude faster than in V2.01. See changes introduced by SVM^struct V3.00 for a general properties of the new structural SVM training algorithm.
• New IO routines that are faster for reading large data files.
• Source code for SVM^cfg V2.01.

V2.00 - V2.01

• Fixed bug that lead to ignoring additional rule features in svm_cfg_classify .

References

[1] I. Tsochantaridis, T. Hofmann, T. Joachims, and Y. Altun. Support Vector Learning for Interdependent and Structured Output Spaces, ICML, 2004. [Postscript] [PDF]

[2] M. Johnson, PCFG models of linguistic tree representations. Computational Linguistics, 1999.

[3] T. Joachims, Making Large-Scale SVM Learning Practical. Advances in Kernel Methods - Support Vector Learning, B. Schölkopf and C. Burges and A. Smola (ed.), MIT Press, 1999. [Postscript (gz)] [PDF]