svm_struct_api_types.h

SVMcfg: Learns a weighted context free grammar from examples. Training examples (e.g. for natural la

/***********************************************************************/
/*                                                                     */
/*   svm_struct_api.h       (modified for PCFG parsing)                */
/*                                                                     */
/*   Definition of API for attaching implementing SVM learning of      */
/*   structures (e.g. parsing, multi-label classification, HMM)        */ 
/*                                                                     */
/*   Author: Thorsten Joachims                                         */
/*   Date: 12.07.04                                                    */
/*                                                                     */
/*   Copyright (c) 2004  Thorsten Joachims - All rights reserved       */
/*                                                                     */
/*   This software is available for non-commercial use only. It must   */
/*   not be modified and distributed without prior permission of the   */
/*   author. The author is not responsible for implications from the   */
/*   use of this software.                                             */
/*                                                                     */
/***********************************************************************/

#ifndef svm_struct_api_types
#define svm_struct_api_types

#define MAXFEAT         28   /* maximum number of features in a rule */

#include "tree.h"
#include "vindex.h"
#include "grammar.h"
#include "svm_light/svm_common.h"
#include "svm_light/svm_learn.h"

#define INST_NAME          "Context-Free Grammar"
#define INST_VERSION       "V3.00"
#define INST_VERSION_DATE  "23.10.06"

/* default precision for solving the optimization problem */
# define DEFAULT_EPS         0.1 
/* default loss rescaling method: 1=slack_rescaling, 2=margin_rescaling */
# define DEFAULT_RESCALING   2
/* default loss function: */
# define DEFAULT_LOSS_FCT    0
/* default optimization algorithm to use: */
# define DEFAULT_ALG_TYPE    4
/* store Psi(x,y) once instead of recomputing it every time: */
# define USE_FYCACHE         1

typedef struct pattern {
  /* this defines the x-part of a training example, e.g. the structure
     for storing a natural language sentence in NLP parsing */
  struct vindex sentence;
  si_t          si;
} PATTERN;

typedef struct label {
  /* this defines the y-part (the label) of a training example,
     e.g. the parse tree of the corresponding sentence. */
  tree parse;
  si_t si;
  double prob;
  double loss;
} LABEL;

typedef struct struct_learn_parm {
  double epsilon;              /* precision for which to solve
				  quadratic program */
  double newconstretrain;      /* number of new constraints to
				  accumulate before recomputing the QP
				  solution */
  int    ccache_size;          /* maximum number of constraints to
				  cache for each example (used in w=4
				  algorithm) */
  double C;                    /* trade-off between margin and loss */
  char   custom_argv[20][300]; /* string set with the -u command line option */
  int    custom_argc;          /* number of -u command line options */
  int    slack_norm;           /* norm to use in objective function
                                  for slack variables; 1 -> L1-norm, 
				  2 -> L2-norm */
  int    loss_type;            /* selected loss function from -r
				  command line option. Select between
				  slack rescaling (1) and margin
				  rescaling (2) */
  int    loss_function;        /* select between different loss
				  functions via -l command line
				  option */
  /* further parameters that are passed to init_struct_model() */
  int    maxsentlen;
  si_t   si;
  int    parent_annotation;
  int    feat_borders;
  int    feat_parent_span_length;
  int    feat_children_span_length;
  int    feat_diff_children_length;
} STRUCT_LEARN_PARM;

typedef struct structmodel {
  double *w;          /* pointer to the learned weights */
  MODEL  *svm_model;  /* the learned SVM model */
  long   sizePsi;     /* maximum number of weights in w */
  /* other information that is needed for the stuctural model can be
     added here, e.g. the grammar rules for NLP parsing */
  grammar grammar;
  si_t    si;
  vihashl weightid_ht;
  STRUCT_LEARN_PARM *sparm;
} STRUCTMODEL;

typedef struct struct_test_stats {
  /* you can add variables for keeping statistics when evaluating the
     test predictions in svm_struct_classify. This can be used in the
     function eval_prediction and print_struct_testing_stats. */
  long parsed_sentences;
  long test_bracket_sum;
  long parse_bracket_sum;
  long common_bracket_sum;
} STRUCT_TEST_STATS;

#endif