src/relevance.c

   1 /* This file is part of Pazpar2.
   2    Copyright (C) 2006-2013 Index Data
   3
   4 Pazpar2 is free software; you can redistribute it and/or modify it under
   5 the terms of the GNU General Public License as published by the Free
   6 Software Foundation; either version 2, or (at your option) any later
   7 version.
   8
   9 Pazpar2 is distributed in the hope that it will be useful, but WITHOUT ANY
  10 WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11 FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12 for more details.
  13
  14 You should have received a copy of the GNU General Public License
  15 along with this program; if not, write to the Free Software
  16 Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17
  18 */
  19
  20 #if HAVE_CONFIG_H
  21 #include <config.h>
  22 #endif
  23
  24 #include <assert.h>
  25 #include <math.h>
  26 #include <stdlib.h>
  27
  28 #include "relevance.h"
  29 #include "session.h"
  30 #include "client.h"
  31 #include "settings.h"
  32
  33 #ifdef WIN32
  34 #define log2(x) (log(x)/log(2))
  35 #endif
  36
  37 struct relevance
  38 {
  39     int *doc_frequency_vec;
  40     int *term_frequency_vec_tmp;
  41     int *term_pos;
  42     int vec_len;
  43     struct word_entry *entries;
  44     pp2_charset_token_t prt;
  45     int rank_cluster;
  46     double follow_factor;
  47     double lead_decay;
  48     int length_divide;
  49     NMEM nmem;
  50     struct norm_client *norm;
  51 };
  52
  53 struct word_entry {
  54     const char *norm_str;
  55     const char *display_str;
  56     int termno;
  57     char *ccl_field;
  58     struct word_entry *next;
  59 };
  60
  61 // Structure to keep data for norm_client scores from one client
  62 struct norm_client
  63 {
  64     int num; // number of the client
  65     float max;
  66     float min;
  67     int count;
  68     const char *native_score;
  69     int scorefield;
  70     float a,b; // Rn = a*R + b
  71     struct client *client;
  72     struct norm_client *next;
  73     struct norm_record *records;
  74 };
  75
  76 const int scorefield_none = -1;  // Do not normalize anything, use tf/idf as is
  77   // This is the old behavior, and the default
  78 const int scorefield_internal = -2;  // use our tf/idf, but normalize it
  79 const int scorefield_position = -3;  // fake a score based on the position
  80
  81 // A structure for each (sub)record. There is one list for each client
  82 struct norm_record
  83 {
  84     struct record *record;
  85     float score;
  86     struct record_cluster *clust;
  87     struct norm_record *next;
  88 };
  89
  90 // Find the norm_client entry for this client, or create one if not there
  91 struct norm_client *findnorm( struct relevance *rel, struct client* client)
  92 {
  93     struct norm_client *n = rel->norm;
  94     struct session_database *sdb;
  95     while (n) {
  96         if (n->client == client )
  97             return n;
  98         n = n->next;
  99     }
 100     n = nmem_malloc(rel->nmem, sizeof(struct norm_client) );
 101     if ( rel->norm )
 102         n->num = rel->norm->num +1;
 103     else
 104         n->num = 1;
 105     n->count = 0;
 106     n->max = 0.0;
 107     n->min = 0.0;
 108     n->client = client;
 109     n->next = rel->norm;
 110     rel->norm = n;
 111     sdb = client_get_database(client);
 112     n->native_score = session_setting_oneval(sdb, PZ_NATIVE_SCORE);
 113     n->records = 0;
 114     n->scorefield = scorefield_none;
 115     yaz_log(YLOG_LOG,"Normalizing: Client %d uses '%s'", n->num, n->native_score );
 116     if ( ! n->native_score  || ! *n->native_score )  // not specified
 117         n->scorefield = scorefield_none;
 118     else if ( strcmp(n->native_score,"position") == 0 )
 119         n->scorefield = scorefield_position;
 120     else if ( strcmp(n->native_score,"internal") == 0 )
 121         n->scorefield = scorefield_internal;
 122     else
 123     { // Get the field index for the score
 124         struct session *se = client_get_session(client);
 125         n->scorefield = conf_service_metadata_field_id(se->service, n->native_score);
 126     }
 127     yaz_log(YLOG_LOG,"Normalizing: Client %d uses '%s' = %d",
 128             n->num, n->native_score, n->scorefield );
 129     return n;
 130 }
 131
 132
 133 // Add a record in the list for that client, for normalizing later
 134 static void setup_norm_record( struct relevance *rel,  struct record_cluster *clust)
 135 {
 136     struct record *record;
 137     for (record = clust->records; record; record = record->next)
 138     {
 139         struct norm_client *norm = findnorm(rel, record->client);
 140         struct norm_record *rp;
 141         if ( norm->scorefield == scorefield_none)
 142             break;  // not interested in normalizing this client
 143         rp = nmem_malloc(rel->nmem, sizeof(struct norm_record) );
 144         norm->count ++;
 145         rp->next = norm->records;
 146         norm->records = rp;
 147         rp->clust = clust;
 148         rp->record = record;
 149         if ( norm->scorefield == scorefield_position )
 150             rp->score = 1.0 / record->position;
 151         else if ( norm->scorefield == scorefield_internal )
 152             rp->score = clust->relevance_score; // the tf/idf for the whole cluster
 153               // TODO - Get them for each record, merge later!
 154         else
 155         {
 156             struct record_metadata *md = record->metadata[norm->scorefield];
 157             rp->score = md->data.fnumber;
 158         }
 159         yaz_log(YLOG_LOG,"Got score for %d/%d : %f ",
 160                 norm->num, record->position, rp->score );
 161         if ( norm->count == 1 )
 162         {
 163             norm->max = rp->score;
 164             norm->min = rp->score;
 165         } else {
 166             if ( rp->score > norm->max )
 167                 norm->max = rp->score;
 168             if ( rp->score < norm->min && abs(rp->score) < 1e-6 )
 169                 norm->min = rp->score;  // skip zeroes
 170         }
 171     }
 172 }
 173
 174 // Calculate the squared sum of residuals, that is the difference from
 175 // normalized values to the target curve, which is 1/n
 176 static double squaresum( struct norm_record *rp, double a, double b)
 177 {
 178     double sum = 0.0;
 179     for ( ; rp; rp = rp->next )
 180     {
 181         double target = 1.0 / rp->record->position;
 182         double normscore = rp->score * a + b;
 183         double diff = target - normscore;
 184         sum += diff * diff;
 185     }
 186     return sum;
 187 }
 188
 189 // For each client, normalize scores
 190 static void normalize_scores(struct relevance *rel)
 191 {
 192     const int maxiterations = 1000;
 193     const double enough = 1000.0;  // sets the number of decimals we are happy with
 194     const double stepchange = 0.5; // reduction of the step size when finding middle
 195       // 0.5 sems to be magical, much better than 0.4 or 0.6
 196     struct norm_client *norm;
 197     for ( norm = rel->norm; norm; norm = norm->next )
 198     {
 199         yaz_log(YLOG_LOG,"Normalizing client %d: scorefield=%d count=%d range=%f %f",
 200                 norm->num, norm->scorefield, norm->count, norm->min, norm->max);
 201         norm->a = 1.0; // default normalizing factors, no change
 202         norm->b = 0.0;
 203         if ( norm->scorefield != scorefield_none &&
 204              norm->scorefield != scorefield_position )
 205         { // have something to normalize
 206             double range = norm->max - norm->min;
 207             int it = 0;
 208             double a,b;  // params to optimize
 209             double as,bs; // step sizes
 210             double chi;
 211             char *branch = "?";
 212             // initial guesses for the parameters
 213             if ( range < 1e-6 ) // practically zero
 214                 range = norm->max;
 215             a = 1.0 / range;
 216             b = abs(norm->min);
 217             as = a / 10;
 218             bs = b / 10;
 219             chi = squaresum( norm->records, a,b);
 220             while (it++ < maxiterations)  // safeguard against things not converging
 221             {
 222                 double aplus = squaresum(norm->records, a+as, b);
 223                 double aminus= squaresum(norm->records, a-as, b);
 224                 double bplus = squaresum(norm->records, a, b+bs);
 225                 double bminus= squaresum(norm->records, a, b-bs);
 226                 double prevchi = chi;
 227                 if ( aplus < chi && aplus < aminus && aplus < bplus && aplus < bminus)
 228                 {
 229                     a = a + as;
 230                     chi = aplus;
 231                     as = as * (1.0 + stepchange);
 232                     branch = "aplus ";
 233                 }
 234                 else if ( aminus < chi && aminus < aplus && aminus < bplus && aminus < bminus)
 235                 {
 236                     a = a - as;
 237                     chi = aminus;
 238                     as = as * (1.0 + stepchange);
 239                     branch = "aminus";
 240                 }
 241                 else if ( bplus < chi && bplus < aplus && bplus < aminus && bplus < bminus)
 242                 {
 243                     b = b + bs;
 244                     chi = bplus;
 245                     bs = bs * (1.0 + stepchange);
 246                     branch = "bplus ";
 247                 }
 248                 else if ( bminus < chi && bminus < aplus && bminus < bplus && bminus < aminus)
 249                 {
 250                     b = b - bs;
 251                     chi = bminus;
 252                     branch = "bminus";
 253                     bs = bs * (1.0+stepchange);
 254                 }
 255                 else
 256                 { // a,b is the best so far, adjust one step size
 257                   // which one? The one that has the greatest effect to chi
 258                   // That is, the average of plus and minus is further away from chi
 259                     double adif = 0.5 * ( aplus + aminus ) - prevchi;
 260                     double bdif = 0.5 * ( bplus + bminus ) - prevchi;
 261                     if ( fabs(adif) > fabs(bdif) )
 262                     {
 263                         as = as * ( 1.0 - stepchange);
 264                         branch = "step a";
 265                     }
 266                     else
 267                     {
 268                         bs = bs * ( 1.0 - stepchange);
 269                         branch = "step b";
 270                     }
 271                 }
 272                 yaz_log(YLOG_LOG,"Fitting %s it=%d: a=%f %f  b=%f %f  chi=%f ap=%f am=%f, bp=%f bm=%f p=%f",
 273                     branch, it, a, as, b, bs, chi,
 274                     aplus, aminus, bplus, bminus, prevchi );
 275                 norm->a = a;
 276                 norm->b = b;
 277                 if ( fabs(as) * enough < fabs(a) &&
 278                      fabs(bs) * enough < fabs(b) ) {
 279                     break;  // not changing much any more
 280
 281                 }
 282             }
 283             yaz_log(YLOG_LOG,"Fitting done: it=%d: a=%f / %f  b=%f / %f  chi = %f",
 284                         it-1, a, as, b, bs, chi );
 285             yaz_log(YLOG_LOG,"  a: %f < %f %d",
 286                     fabs(as)*enough, fabs(a), (fabs(as) * enough < fabs(a)) );
 287             yaz_log(YLOG_LOG,"  b: %f < %f %d",
 288                     fabs(bs)*enough, fabs(b), (fabs(bs) * enough < fabs(b)) );
 289         }
 290
 291         if ( norm->scorefield != scorefield_none )
 292         { // distribute the normalized scores to the records
 293             struct norm_record *nr = norm->records;
 294             for ( ; nr ; nr = nr->next ) {
 295                 double r = nr->score;
 296                 r = norm->a * r + norm -> b;
 297                 nr->clust->relevance_score = 10000 * r;
 298                 yaz_log(YLOG_LOG,"Normalized %f * %f + %f = %f",
 299                         nr->score, norm->a, norm->b, r );
 300                 // TODO - This keeps overwriting the cluster score in random order!
 301                 // Need to merge results better
 302             }
 303
 304         }
 305
 306     } // client loop
 307 }
 308
 309
 310 static struct word_entry *word_entry_match(struct relevance *r,
 311                                            const char *norm_str,
 312                                            const char *rank, int *weight)
 313 {
 314     int i = 1;
 315     struct word_entry *entries = r->entries;
 316     for (; entries; entries = entries->next, i++)
 317     {
 318         if (*norm_str && !strcmp(norm_str, entries->norm_str))
 319         {
 320             const char *cp = 0;
 321             int no_read = 0;
 322             sscanf(rank, "%d%n", weight, &no_read);
 323             rank += no_read;
 324             while (*rank == ' ')
 325                 rank++;
 326             if (no_read > 0 && (cp = strchr(rank, ' ')))
 327             {
 328                 if ((cp - rank) == strlen(entries->ccl_field) &&
 329                     memcmp(entries->ccl_field, rank, cp - rank) == 0)
 330                     *weight = atoi(cp + 1);
 331             }
 332             return entries;
 333         }
 334     }
 335     return 0;
 336 }
 337
 338 int relevance_snippet(struct relevance *r,
 339                       const char *words, const char *name,
 340                       WRBUF w_snippet)
 341 {
 342     int no = 0;
 343     const char *norm_str;
 344     int highlight = 0;
 345
 346     pp2_charset_token_first(r->prt, words, 0);
 347     while ((norm_str = pp2_charset_token_next(r->prt)))
 348     {
 349         size_t org_start, org_len;
 350         struct word_entry *entries = r->entries;
 351         int i;
 352
 353         pp2_get_org(r->prt, &org_start, &org_len);
 354         for (; entries; entries = entries->next, i++)
 355         {
 356             if (*norm_str && !strcmp(norm_str, entries->norm_str))
 357                 break;
 358         }
 359         if (entries)
 360         {
 361             if (!highlight)
 362             {
 363                 highlight = 1;
 364                 wrbuf_puts(w_snippet, "<match>");
 365                 no++;
 366             }
 367         }
 368         else
 369         {
 370             if (highlight)
 371             {
 372                 highlight = 0;
 373                 wrbuf_puts(w_snippet, "</match>");
 374             }
 375         }
 376         wrbuf_xmlputs_n(w_snippet, words + org_start, org_len);
 377     }
 378     if (highlight)
 379         wrbuf_puts(w_snippet, "</match>");
 380     if (no)
 381     {
 382         yaz_log(YLOG_DEBUG, "SNIPPET match: %s", wrbuf_cstr(w_snippet));
 383     }
 384     return no;
 385 }
 386
 387 void relevance_countwords(struct relevance *r, struct record_cluster *cluster,
 388                           const char *words, const char *rank,
 389                           const char *name)
 390 {
 391     int *w = r->term_frequency_vec_tmp;
 392     const char *norm_str;
 393     int i, length = 0;
 394     double lead_decay = r->lead_decay;
 395     struct word_entry *e;
 396     WRBUF wr = cluster->relevance_explain1;
 397     int printed_about_field = 0;
 398
 399     pp2_charset_token_first(r->prt, words, 0);
 400     for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
 401     {
 402         w[i] = 0;
 403         r->term_pos[i] = 0;
 404     }
 405
 406     assert(rank);
 407     while ((norm_str = pp2_charset_token_next(r->prt)))
 408     {
 409         int local_weight = 0;
 410         e = word_entry_match(r, norm_str, rank, &local_weight);
 411         if (e)
 412         {
 413             int res = e->termno;
 414             int j;
 415
 416             if (!printed_about_field)
 417             {
 418                 printed_about_field = 1;
 419                 wrbuf_printf(wr, "field=%s content=", name);
 420                 if (strlen(words) > 50)
 421                 {
 422                     wrbuf_xmlputs_n(wr, words, 49);
 423                     wrbuf_puts(wr, " ...");
 424                 }
 425                 else
 426                     wrbuf_xmlputs(wr, words);
 427                 wrbuf_puts(wr, ";\n");
 428             }
 429             assert(res < r->vec_len);
 430             w[res] += local_weight / (1 + log2(1 + lead_decay * length));
 431             wrbuf_printf(wr, "%s: w[%d] += w(%d) / "
 432                          "(1+log2(1+lead_decay(%f) * length(%d)));\n",
 433                          e->display_str, res, local_weight, lead_decay, length);
 434             j = res - 1;
 435             if (j > 0 && r->term_pos[j])
 436             {
 437                 int d = length + 1 - r->term_pos[j];
 438                 wrbuf_printf(wr, "%s: w[%d] += w[%d](%d) * follow(%f) / "
 439                              "(1+log2(d(%d));\n",
 440                              e->display_str, res, res, w[res],
 441                              r->follow_factor, d);
 442                 w[res] += w[res] * r->follow_factor / (1 + log2(d));
 443             }
 444             for (j = 0; j < r->vec_len; j++)
 445                 r->term_pos[j] = j < res ? 0 : length + 1;
 446         }
 447         length++;
 448     }
 449
 450     for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
 451     {
 452         if (length == 0 || w[i] == 0)
 453             continue;
 454         wrbuf_printf(wr, "%s: tf[%d] += w[%d](%d)", e->display_str, i, i, w[i]);
 455         switch (r->length_divide)
 456         {
 457         case 0:
 458             cluster->term_frequency_vecf[i] += (double) w[i];
 459             break;
 460         case 1:
 461             wrbuf_printf(wr, " / log2(1+length(%d))", length);
 462             cluster->term_frequency_vecf[i] +=
 463                 (double) w[i] / log2(1 + length);
 464             break;
 465         case 2:
 466             wrbuf_printf(wr, " / length(%d)", length);
 467             cluster->term_frequency_vecf[i] += (double) w[i] / length;
 468         }
 469         cluster->term_frequency_vec[i] += w[i];
 470         wrbuf_printf(wr, " (%f);\n", cluster->term_frequency_vecf[i]);
 471     }
 472
 473     cluster->term_frequency_vec[0] += length;
 474 }
 475
 476 static void pull_terms(struct relevance *res, struct ccl_rpn_node *n)
 477 {
 478     char **words;
 479     int numwords;
 480     char *ccl_field;
 481     int i;
 482
 483     switch (n->kind)
 484     {
 485     case CCL_RPN_AND:
 486     case CCL_RPN_OR:
 487     case CCL_RPN_NOT:
 488     case CCL_RPN_PROX:
 489         pull_terms(res, n->u.p[0]);
 490         pull_terms(res, n->u.p[1]);
 491         break;
 492     case CCL_RPN_TERM:
 493         nmem_strsplit(res->nmem, " ", n->u.t.term, &words, &numwords);
 494         for (i = 0; i < numwords; i++)
 495         {
 496             const char *norm_str;
 497
 498             ccl_field = nmem_strdup_null(res->nmem, n->u.t.qual);
 499
 500             pp2_charset_token_first(res->prt, words[i], 0);
 501             while ((norm_str = pp2_charset_token_next(res->prt)))
 502             {
 503                 struct word_entry **e = &res->entries;
 504                 while (*e)
 505                     e = &(*e)->next;
 506                 *e = nmem_malloc(res->nmem, sizeof(**e));
 507                 (*e)->norm_str = nmem_strdup(res->nmem, norm_str);
 508                 (*e)->ccl_field = ccl_field;
 509                 (*e)->termno = res->vec_len++;
 510                 (*e)->display_str = nmem_strdup(res->nmem, words[i]);
 511                 (*e)->next = 0;
 512             }
 513         }
 514         break;
 515     default:
 516         break;
 517     }
 518 }
 519 void relevance_clear(struct relevance *r)
 520 {
 521     if (r)
 522     {
 523         int i;
 524         for (i = 0; i < r->vec_len; i++)
 525             r->doc_frequency_vec[i] = 0;
 526     }
 527 }
 528
 529 struct relevance *relevance_create_ccl(pp2_charset_fact_t pft,
 530                                        struct ccl_rpn_node *query,
 531                                        int rank_cluster,
 532                                        double follow_factor, double lead_decay,
 533                                        int length_divide)
 534 {
 535     NMEM nmem = nmem_create();
 536     struct relevance *res = nmem_malloc(nmem, sizeof(*res));
 537
 538     res->nmem = nmem;
 539     res->entries = 0;
 540     res->vec_len = 1;
 541     res->rank_cluster = rank_cluster;
 542     res->follow_factor = follow_factor;
 543     res->lead_decay = lead_decay;
 544     res->length_divide = length_divide;
 545     res->norm = 0;
 546     res->prt = pp2_charset_token_create(pft, "relevance");
 547
 548     pull_terms(res, query);
 549
 550     res->doc_frequency_vec = nmem_malloc(nmem, res->vec_len * sizeof(int));
 551
 552     // worker array
 553     res->term_frequency_vec_tmp =
 554         nmem_malloc(res->nmem,
 555                     res->vec_len * sizeof(*res->term_frequency_vec_tmp));
 556
 557     res->term_pos =
 558         nmem_malloc(res->nmem, res->vec_len * sizeof(*res->term_pos));
 559
 560     relevance_clear(res);
 561     return res;
 562 }
 563
 564 void relevance_destroy(struct relevance **rp)
 565 {
 566     if (*rp)
 567     {
 568         pp2_charset_token_destroy((*rp)->prt);
 569         nmem_destroy((*rp)->nmem);
 570         *rp = 0;
 571     }
 572 }
 573
 574 void relevance_mergerec(struct relevance *r, struct record_cluster *dst,
 575                         const struct record_cluster *src)
 576 {
 577     int i;
 578
 579     for (i = 0; i < r->vec_len; i++)
 580         dst->term_frequency_vec[i] += src->term_frequency_vec[i];
 581
 582     for (i = 0; i < r->vec_len; i++)
 583         dst->term_frequency_vecf[i] += src->term_frequency_vecf[i];
 584 }
 585
 586 void relevance_newrec(struct relevance *r, struct record_cluster *rec)
 587 {
 588     int i;
 589
 590     // term frequency [1,..] . [0] is total length of all fields
 591     rec->term_frequency_vec =
 592         nmem_malloc(r->nmem,
 593                     r->vec_len * sizeof(*rec->term_frequency_vec));
 594     for (i = 0; i < r->vec_len; i++)
 595         rec->term_frequency_vec[i] = 0;
 596
 597     // term frequency divided by length of field [1,...]
 598     rec->term_frequency_vecf =
 599         nmem_malloc(r->nmem,
 600                     r->vec_len * sizeof(*rec->term_frequency_vecf));
 601     for (i = 0; i < r->vec_len; i++)
 602         rec->term_frequency_vecf[i] = 0.0;
 603 }
 604
 605 void relevance_donerecord(struct relevance *r, struct record_cluster *cluster)
 606 {
 607     int i;
 608
 609     for (i = 1; i < r->vec_len; i++)
 610         if (cluster->term_frequency_vec[i] > 0)
 611             r->doc_frequency_vec[i]++;
 612
 613     r->doc_frequency_vec[0]++;
 614 }
 615
 616
 617
 618 // Prepare for a relevance-sorted read
 619 void relevance_prepare_read(struct relevance *rel, struct reclist *reclist)
 620 {
 621     int i;
 622     float *idfvec = xmalloc(rel->vec_len * sizeof(float));
 623
 624     reclist_enter(reclist);
 625
 626     // Calculate document frequency vector for each term.
 627     for (i = 1; i < rel->vec_len; i++)
 628     {
 629         if (!rel->doc_frequency_vec[i])
 630             idfvec[i] = 0;
 631         else
 632         {
 633             /* add one to nominator idf(t,D) to ensure a value > 0 */
 634             idfvec[i] = log((float) (1 + rel->doc_frequency_vec[0]) /
 635                             rel->doc_frequency_vec[i]);
 636         }
 637     }
 638     // Calculate relevance for each document
 639     while (1)
 640     {
 641         int relevance = 0;
 642         WRBUF w;
 643         struct word_entry *e = rel->entries;
 644         struct record_cluster *rec = reclist_read_record(reclist);
 645         if (!rec)
 646             break;
 647         w = rec->relevance_explain2;
 648         wrbuf_rewind(w);
 649         wrbuf_puts(w, "relevance = 0;\n");
 650         for (i = 1; i < rel->vec_len; i++)
 651         {
 652             float termfreq = (float) rec->term_frequency_vecf[i];
 653             int add = 100000 * termfreq * idfvec[i];
 654
 655             wrbuf_printf(w, "idf[%d] = log(((1 + total(%d))/termoccur(%d));\n",
 656                          i, rel->doc_frequency_vec[0],
 657                          rel->doc_frequency_vec[i]);
 658             wrbuf_printf(w, "%s: relevance += 100000 * tf[%d](%f) * "
 659                          "idf[%d](%f) (%d);\n",
 660                          e->display_str, i, termfreq, i, idfvec[i], add);
 661             relevance += add;
 662             e = e->next;
 663         }
 664         if (!rel->rank_cluster)
 665         {
 666             struct record *record;
 667             int cluster_size = 0;
 668
 669             for (record = rec->records; record; record = record->next)
 670                 cluster_size++;
 671
 672             wrbuf_printf(w, "score = relevance(%d)/cluster_size(%d);\n",
 673                          relevance, cluster_size);
 674             relevance /= cluster_size;
 675         }
 676         else
 677         {
 678             wrbuf_printf(w, "score = relevance(%d);\n", relevance);
 679         }
 680         rec->relevance_score = relevance;
 681
 682         // Build the normalizing structures
 683         // List of (sub)records for each target
 684         setup_norm_record( rel, rec );
 685
 686         // TODO - Loop again, merge individual record scores into clusters
 687         // Can I reset the reclist, or can I leave and enter without race conditions?
 688
 689     } // cluster loop
 690
 691     normalize_scores(rel);
 692
 693     reclist_leave(reclist);
 694     xfree(idfvec);
 695
 696 }
 697
 698 /*
 699  * Local variables:
 700  * c-basic-offset: 4
 701  * c-file-style: "Stroustrup"
 702  * indent-tabs-mode: nil
 703  * End:
 704  * vim: shiftwidth=4 tabstop=8 expandtab
 705  */
 706