Blame view
tools/sctk-2.4.10/src/sclite/sentmcn.c
17.1 KB
8dcb6dfcb
first commit
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 |
/********************************************************************/
/* */
/* FILE: McNemar_sent.c */
/* WRITTEN BY: Jonathan G. Fiscus */
/* DATE: May 31 1989 */
/* NATIONAL INSTITUTE OF STANDARDS AND TECHNOLOGY */
/* SPEECH RECOGNITION GROUP */
/* */
/* USAGE: This program, performs a McNemar test on the */
/* list of SYS_ALIGN structures */
/* The test statistics models the CHI SQUARE dist */
/* and is calculated from the McNemar matrix: */
/* */
/* McNemar matrix: */
/* sys1 C: denotes a corr sent */
/* C E E: denotes a error sent */
/* ------------- TC: sum of C */
/* C | a | b | TC TE: sum of E */
/* sys2 |-----+-----| */
/* E | c | d | TE element definitions: */
/* ------------- a: num corr by both */
/* TC TE systems */
/* c: num corr by sys1 */
/* but not by sys2 */
/* b: num corr by sys2 */
/* but not by sys1 */
/* d: num errored on by */
/* both systems */
/* Test Statistic formula: */
/* */
/* */
/* TS = Binomial(MIN(c,b),c+b,0.5) */
/* */
/* */
/********************************************************************/
#include "sctk.h"
static int compute_McNemar(int **table, char *treat1_str, char *treat2_str, int verbosely, FILE *fp, double *conf, double alpha);
static void print_compare_matrix_for_sent_M(SCORES *scor[], int nscor, int **winner, double **conf, char *tname, char *matrix_name, FILE *fp);
/********************************************************************/
/* this procedure does all the system comparisons then it */
/* prints a report to stdout */
/********************************************************************/
void McNemar_sent(SCORES *scor[], int nscor, int ***out_winner, char *testname, int print_results, int verbose, char *outroot, int feedback, double ***out_conf)
{
int comp1, comp2, **winner, result;
double **conf;
FILE *fp = stdout;
if (print_results || verbose){
char *f = rsprintf("%s.mcn",outroot);
if ((fp=(strcmp(outroot,"-") == 0) ? stdout : fopen(f,"w")) ==
(FILE *)0){
fprintf(stderr,"Warning: Open of %s for write failed. "
"Using stdout instead.
",f);
fp = stdout;
}
if (feedback >= 1) printf(" Output written to '%s'
",f);
}
alloc_2dimZ(winner,nscor,nscor,int,NO_DIFF);
*out_winner = winner;
alloc_2dimZ(conf,nscor,nscor,double,0.0);
*out_conf = conf;
for (comp1=0; comp1 <(nscor-1); comp1++)
for (comp2=comp1+1; comp2<nscor; comp2++){
result = do_McNemar_by_sent(scor[comp1],scor[comp2],verbose,fp,
&(conf[comp1][comp2]));
winner[comp1][comp2] = result;
}
if (print_results){
if (verbose) form_feed(fp);
print_compare_matrix_for_sent_M(scor, nscor, winner, conf, testname,
"COMPARISON MATRIX: McNEMAR\'S TEST ON CORRECT SENTENCES FOR THE TEST:", fp);
if (fp == stdout) form_feed(fp);
}
if (fp != stdout) fclose(fp);
}
/********************************************************************/
/* using the COUNT structure, calculate the matrix of the McNemar */
/* test then perform the test */
/********************************************************************/
int do_McNemar_by_sent(SCORES *sys1, SCORES *sys2, int verbose, FILE *fp, double *conf)
{
int ans, spk1, spk2, snt1, snt2, e1, e2;
PATH *cp;
int **table=NULL, nw;
int foundMatchSent;
alloc_2dimZ(table,2,2,int,0);
for (spk1=0;spk1 < sys1->num_grp; spk1++){ /* for all speaker sys1 */
/**** find the matching speaker */
for (spk2=0;spk2 < sys2->num_grp; spk2++)
if (strcmp(sys1->grp[spk1].name, sys2->grp[spk2].name) == 0)
break;
/**** the the speakers match, start on the sentences */
if (spk2 != sys2->num_grp){
/**** for all sents in sys1,spkr1 */
for (snt1 = 0; snt1 < sys1->grp[spk1].num_path; snt1++){
/**** for all sents in sys2,spkr2 */
foundMatchSent = 0;
for (snt2 = 0; snt2 < sys2->grp[spk2].num_path; snt2++){
/**** if the sentences are the same, compare them */
if(strcmp(sys1->grp[spk1].path[snt1]->id,
sys2->grp[spk2].path[snt2]->id) == 0){
e1 = e2 = 0;
cp = sys1->grp[spk1].path[snt1];
for (nw=0; nw < cp->num; nw++)
if (cp->pset[nw].eval != P_CORR) {
e1 = 1;
break;
}
cp = sys2->grp[spk2].path[snt2];
for (nw=0; nw < cp->num; nw++)
if (cp->pset[nw].eval != P_CORR) {
e2 = 1;
break;
}
table[e1][e2] ++;
foundMatchSent = 1;
}
}
if (! foundMatchSent)
fprintf(stderr,"Warning: Speaker's '%s' path '%s' in system '%s' is not in system '%s'
",
sys1->grp[spk1].name,sys1->grp[spk1].path[snt1]->id,sys1->title,sys2->title );
}
} else {
fprintf(stderr,"Warning: Speaker %s is in system %s but not system %s
",
sys1->grp[spk1].name,sys1->title,sys2->title);
}
}
ans = do_McNemar(table,sys1->title,sys2->title,verbose,fp,conf);
free_2dimarr(table,2,int);
return(ans);
}
/********************************************************************/
/* given the McNemar matrix, come up with an answer. verbose if */
/* desired */
/********************************************************************/
int do_McNemar(int **table, char *name1, char *name2, int verbose, FILE *fp, double *conf)
{
if (verbose){
fprintf(fp,"
");
fprintf(fp," McNemar test results
");
fprintf(fp," ====================
");
fprintf(fp," %s
",name2);
fprintf(fp," corr incorr
");
fprintf(fp," %15s corr %3d %3d
",name1,
table[0][0],table[0][1]);
fprintf(fp,"\t incorr %3d %3d
",
table[1][0],table[1][1]);
}
if (((table[0][1] == 0) && (table[1][0] == 0)) ||
(table[0][1] == table[1][0])){
/* if both of these feilds are 0, or equal, then there is no */
/* difference so say so */
if (verbose){
fprintf(fp,"
\t\tSUMMARY:
\t\t-------
");
fprintf(fp,"
\tThe two totals for utterances missed by either test results
");
fprintf(fp,"\tare both zero, therfore there is no significant difference
");
fprintf(fp,"\tbetween the two tests!
");
}
*conf = 1.00;
return(NO_DIFF);
}
else
return(compute_McNemar(table,name1,name2,verbose,fp,conf,0.05));
}
#ifdef UNUSED
/********************************************************************/
/* this program calculates the test statistic corresponding to a */
/* CHI SQUARED distribution. the formula is at the top of this */
/* file. */
/* references to the 'Peregoy method' are adapted from */
/* Appendix B "The Significance of Bench Mark Test Results" by */
/* Peter Peregoy in American National Standard IAI 1-1987. */
/* Other distributions might be assumed. */
/********************************************************************/
static int perform_peregoy_method(int **table, int verbosely,FILE *fp, double *conf)
{
int i;
double ts_per;
extern double sqrt(double);
/* Perogoy method */
ts_per = ((fabs((double)(table[0][1]-table[1][0]))-1)*
(fabs((double)(table[0][1]-table[1][0]))-1)) /
(table[0][1]+table[1][0]);
/* compute the confidence measure */
*conf = X2.per[MIN_X2_PER] / 100.0;
for (i=MIN_X2_PER;i<MAX_X2_PER;i++)
if (X2.df[DF1].level[i] < fabs(ts_per) &&
X2.df[DF1].level[i+1] > fabs(ts_per)){
*conf = X2.per[i] / 100.0;
break;
}
if (X2.df[DF1].level[MAX_X2_PER] < fabs(ts_per))
*conf = X2.per[MAX_X2_PER] / 100.0;
if (verbosely){
fprintf(fp,"
");
fprintf(fp,"%30s Reject if
","");
fprintf(fp,"%27sX > X2 of %s %s (%2.3f)
", "",
X2.per_str[GEN_X2_PER],
X2.df[DF1].str,
X2.df[DF1].level[GEN_X2_PER]);
fprintf(fp,"
");
fprintf(fp,"%30s X = %2.3f
","",ts_per);
fprintf(fp,"
\t\tSUMMARY:
\t\t-------
");
}
if (fabs(ts_per) > X2.df[DF1].level[GEN_X2_PER]){
if (verbosely){
fprintf(fp,"\tPeregoy's method shows that, with %s confidence, the
",
X2.neg_per_str[GEN_X2_PER]),
fprintf(fp,"\t2 recognition systems are significantly different.
");
fprintf(fp,"
");
fprintf(fp,"\tFurther, the probablity of there being a difference is
");
for (i=GEN_X2_PER;i<MAX_X2_PER;i++)
if (fabs(ts_per) < X2.df[DF1].level[i+1])
break;
if (i==MAX_X2_PER)
fprintf(fp,"\tgreater that %s.
",X2.neg_per_str[i]);
else
fprintf(fp,"\tbetween %s to %s.
",X2.neg_per_str[i],
X2.neg_per_str[i+1]);
fprintf(fp,"
");
}
if (table[1][0] < table[0][1])
return(TEST_DIFF * (-1)); /* invert because sys1 was better */
else
return(TEST_DIFF);
}
else{
if (verbosely){
fprintf(fp,"\tPeregoy's method shows that at the %s confidence
",
X2.neg_per_str[GEN_X2_PER]),
fprintf(fp,"\tinterval, the 2 recognition systems are not significantly
");
fprintf(fp,"\tdifferent.
");
fprintf(fp,"\tFurther, the probablity of there being a difference is
");
for (i=GEN_X2_PER;i>MIN_X2_PER;i--)
if (fabs(ts_per) > X2.df[DF1].level[i-1])
break;
if (i==MIN_X2_PER)
fprintf(fp,"\tless than %s.
",X2.neg_per_str[i]);
else
fprintf(fp,"\tbetween %s to %s.
",X2.neg_per_str[i-1],
X2.neg_per_str[i]);
fprintf(fp,"
");
}
return(NO_DIFF);
}
}
#endif
/********************************************************************/
/* this procedure is called to print out a comparison matrix for */
/* any comparison function. */
/********************************************************************/
static void print_compare_matrix_for_sent_M(SCORES *scor[], int nscor, int **winner, double **conf, char *tname, char *matrix_name,FILE *fp)
{
char fmt[50];
int i,j,sys,spkr,snt, *corr_arr;
alloc_singarr(corr_arr,nscor,int);
/* calc the number of correct for each system */
for (sys=0;sys<nscor;sys++){
corr_arr[sys] = 0;
for (spkr=0; spkr<scor[sys]->num_grp; spkr++)
for (snt=0; snt<scor[sys]->grp[spkr].num_path; snt++)
corr_arr[sys] += scor[sys]->grp[spkr].num_path -
scor[sys]->grp[spkr].serr;
}
strcpy(fmt,"c");
for (i=0; i<nscor; i++) strcat(fmt,"|c");
Desc_erase();
Desc_set_page_center(SCREEN_WIDTH);
Desc_add_row_values("c",matrix_name);
Desc_add_row_separation(' ',AFTER_ROW);
Desc_add_row_separation('-',AFTER_ROW);
Desc_add_row_values("c",tname);
Desc_set_iterated_format(fmt);
Desc_set_iterated_value("");
for (i=0; i<nscor; i++)
Desc_set_iterated_value(rsprintf("%s",scor[i]->title));
Desc_flush_iterated_row();
for (i=0; i<nscor; i++) {
Desc_add_row_separation('-',BEFORE_ROW);
Desc_set_iterated_format(fmt);
Desc_set_iterated_value(rsprintf("%s",scor[i]->title));
for (j=0; j<nscor; j++)
if (j > i)
Desc_set_iterated_value(rsprintf("Conf=(%.3f)",(conf[i][j] < 0.001) ? 0.001 : conf[i][j]));
else
Desc_set_iterated_value("");
Desc_flush_iterated_row();
Desc_set_iterated_format(fmt);
Desc_set_iterated_value("");
for (j=0; j<nscor; j++){
char *name="";
if (j > i){
if (winner[i][j] == TEST_DIFF)
name=scor[j]->title;
else if (winner[i][j] == NO_DIFF)
name="same";
else
name=scor[i]->title;
}
Desc_set_iterated_value(name);
}
Desc_flush_iterated_row();
}
Desc_dump_report(1,fp);
free_singarr(corr_arr,int);
}
static int compute_McNemar(int **table, char *treat1_str, char *treat2_str, int verbose, FILE *fp, double *confidence, double alpha)
{
double test_stat, p=0.5;
int decision_cutoff=(-1), i, num_a, num_b;
num_a = table[1][0];
num_b = table[0][1];
/* multiplication by 2 means it's a two-tailed test */
test_stat = 2.0 * compute_acc_binomial(MIN(num_a,num_b),num_a+num_b,p);
*confidence = test_stat;
for (i=0; i< (num_a + num_b); i++)
if (2.0*compute_acc_binomial(i,num_a+num_b,p) < alpha)
decision_cutoff=i;
if (verbose){
fprintf(fp,"The NULL Hypothesis:
");
fprintf(fp," The number unique utterance errors are equal for both systems.
");
fprintf(fp,"
");
fprintf(fp,"Alternate Hypothesis:
");
fprintf(fp," The number of unique utterance errors for both systems are NOT equal.
");
fprintf(fp,"Decision Analysis:
");
fprintf(fp," Assumptions:
");
fprintf(fp," A1: The distibution of unique utterance errors
");
fprintf(fp," follows the binomial distribution for N fair coin tosses.
");
fprintf(fp,"
");
fprintf(fp," Rejection criterion:
");
fprintf(fp," Reject the null hypothesis at the 95%% confidence level based
");
fprintf(fp," on the following critical values table. N is the sum of the
");
fprintf(fp," unique utterance errors for both systems being compared and
");
fprintf(fp," MIN(uue) is the minimum number of unique utterance
");
fprintf(fp," foe either system.
");
/* print a table of critical values */
fprintf(fp," MIN(uue) P(MIN(uue) | N=%3d)
",num_a+num_b);
fprintf(fp," -------- -------------------
");
for (i=0; 2.0*compute_acc_binomial(i-1,num_a+num_b,p) < (MAX(0.25,test_stat)); i++){
double val = 2.0*compute_acc_binomial(i,num_a+num_b,p);
if (val >= 0.0005)
fprintf(fp," %3d %5.3f", i,val);
else
fprintf(fp," %3d - ", i);
if ((val < alpha) && (2.0*compute_acc_binomial(i+1,num_a+num_b,p) > alpha))
fprintf(fp," <--- Null Hypothesis rejection threshold
");
else
fprintf(fp,"
");
}
fprintf(fp,"
");
fprintf(fp," Decision:
");
fprintf(fp," There were MIN(uue)=%d unique utterance errors, the probability of
",
MIN(num_a,num_b));
fprintf(fp," it occuring is %5.3f, therefore the null hypothesis ",test_stat);
if (test_stat < alpha){
fprintf(fp,"is REJECTED
");
fprintf(fp," in favor of the Alternate Hypothesis. Further, %s is the
",
(num_a > num_b) ? treat2_str : treat1_str);
fprintf(fp," better System.
");
} else{
fprintf(fp,"is ACCEPTED
");
fprintf(fp," There is no statistical difference between %s and %s
",treat1_str,treat2_str);
}
form_feed(fp);
}
if (test_stat < alpha){
if (0) fprintf(fp,"Returning Result %d
",TEST_DIFF * ((num_a > num_b) ? 1 : -1));
return(TEST_DIFF * ((num_a > num_b) ? 1 : -1));
}
return(NO_DIFF);
}
|