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             assert(rp->score>0); // ###
 159         }
 160         yaz_log(YLOG_LOG,"Got score for %d/%d : %f ",
 161                 norm->num, record->position, rp->score );
 162         if ( norm->count == 1 )
 163         {
 164             norm->max = rp->score;
 165             norm->min = rp->score;
 166         } else {
 167             if ( rp->score > norm->max )
 168                 norm->max = rp->score;
 169             if ( rp->score < norm->min && abs(rp->score) < 1e-6 )
 170                 norm->min = rp->score;  // skip zeroes
 171         }
 172     }
 173 }
 174
 175 // Calculate the squared sum of residuals, that is the difference from
 176 // normalized values to the target curve, which is 1/n
 177 static double squaresum( struct norm_record *rp, double a, double b)
 178 {
 179     double sum = 0.0;
 180     for ( ; rp; rp = rp->next )
 181     {
 182         double target = 1.0 / rp->record->position;
 183         double normscore = rp->score * a + b;
 184         double diff = target - normscore;
 185         sum += diff * diff;
 186     }
 187     return sum;
 188 }
 189
 190 static void normalize_scores(struct relevance *rel)
 191 {
 192     // For each client, normalize scores
 193     struct norm_client *norm;
 194     for ( norm = rel->norm; norm; norm = norm->next )
 195     {
 196         yaz_log(YLOG_LOG,"Normalizing client %d: scorefield=%d count=%d",
 197                 norm->num, norm->scorefield, norm->count);
 198         norm->a = 1.0; // default normalizing factors, no change
 199         norm->b = 0.0;
 200         if ( norm->scorefield != scorefield_none &&
 201              norm->scorefield != scorefield_position )
 202         { // have something to normalize
 203             double range = norm->max - norm->min;
 204             int it = 0;
 205             double a,b;  // params to optimize
 206             double as,bs; // step sizes
 207             double chi;
 208             // initial guesses for the parameters
 209             if ( range < 1e-6 ) // practically zero
 210                 range = norm->max;
 211             a = 1.0 / range;
 212             b = abs(norm->min);
 213             as = a / 3;
 214             bs = b / 3;
 215             chi = squaresum( norm->records, a,b);
 216             while (it++ < 100)  // safeguard against things not converging
 217             {
 218                 // optimize a
 219                 double plus = squaresum(norm->records, a+as, b);
 220                 double minus= squaresum(norm->records, a-as, b);
 221                 if ( plus < chi && plus < minus )
 222                 {
 223                     a = a + as;
 224                     chi = plus;
 225                 }
 226                 else if ( minus < chi && minus < plus )
 227                 {
 228                     a = a - as;
 229                     chi = minus;
 230                 }
 231                 else
 232                     as = as / 2;
 233                 // optimize b
 234                 plus = squaresum(norm->records, a, b+bs);
 235                 minus= squaresum(norm->records, a, b-bs);
 236                 if ( plus < chi && plus < minus )
 237                 {
 238                     b = b + bs;
 239                     chi = plus;
 240                 }
 241                 else if ( minus < chi && minus < plus )
 242                 {
 243                     b = b - bs;
 244                     chi = minus;
 245                 }
 246                 else
 247                     bs = bs / 2;
 248                 yaz_log(YLOG_LOG,"Fitting it=%d: a=%f / %f  b=%f / %f  chi = %f",
 249                         it, a, as, b, bs, chi );
 250                 norm->a = a;
 251                 norm->b = b;
 252                 if ( abs(as) * 1000.0 < abs(a) &&
 253                      abs(bs) * 1000.0 < abs(b) )
 254                     break;  // not changing much any more
 255             }
 256         }
 257
 258         if ( norm->scorefield != scorefield_none )
 259         { // distribute the normalized scores to the records
 260             struct norm_record *nr = norm->records;
 261             for ( ; nr ; nr = nr->next ) {
 262                 double r = nr->score;
 263                 r = norm->a * r + norm -> b;
 264                 nr->clust->relevance_score = 10000 * r;
 265                 yaz_log(YLOG_LOG,"Normalized %f * %f + %f = %f",
 266                         nr->score, norm->a, norm->b, r );
 267                 // TODO - This keeps overwriting the cluster score in random order!
 268                 // Need to merge results better
 269             }
 270
 271         }
 272
 273     } // client loop
 274 }
 275
 276
 277 static struct word_entry *word_entry_match(struct relevance *r,
 278                                            const char *norm_str,
 279                                            const char *rank, int *weight)
 280 {
 281     int i = 1;
 282     struct word_entry *entries = r->entries;
 283     for (; entries; entries = entries->next, i++)
 284     {
 285         if (*norm_str && !strcmp(norm_str, entries->norm_str))
 286         {
 287             const char *cp = 0;
 288             int no_read = 0;
 289             sscanf(rank, "%d%n", weight, &no_read);
 290             rank += no_read;
 291             while (*rank == ' ')
 292                 rank++;
 293             if (no_read > 0 && (cp = strchr(rank, ' ')))
 294             {
 295                 if ((cp - rank) == strlen(entries->ccl_field) &&
 296                     memcmp(entries->ccl_field, rank, cp - rank) == 0)
 297                     *weight = atoi(cp + 1);
 298             }
 299             return entries;
 300         }
 301     }
 302     return 0;
 303 }
 304
 305 int relevance_snippet(struct relevance *r,
 306                       const char *words, const char *name,
 307                       WRBUF w_snippet)
 308 {
 309     int no = 0;
 310     const char *norm_str;
 311     int highlight = 0;
 312
 313     pp2_charset_token_first(r->prt, words, 0);
 314     while ((norm_str = pp2_charset_token_next(r->prt)))
 315     {
 316         size_t org_start, org_len;
 317         struct word_entry *entries = r->entries;
 318         int i;
 319
 320         pp2_get_org(r->prt, &org_start, &org_len);
 321         for (; entries; entries = entries->next, i++)
 322         {
 323             if (*norm_str && !strcmp(norm_str, entries->norm_str))
 324                 break;
 325         }
 326         if (entries)
 327         {
 328             if (!highlight)
 329             {
 330                 highlight = 1;
 331                 wrbuf_puts(w_snippet, "<match>");
 332                 no++;
 333             }
 334         }
 335         else
 336         {
 337             if (highlight)
 338             {
 339                 highlight = 0;
 340                 wrbuf_puts(w_snippet, "</match>");
 341             }
 342         }
 343         wrbuf_xmlputs_n(w_snippet, words + org_start, org_len);
 344     }
 345     if (highlight)
 346         wrbuf_puts(w_snippet, "</match>");
 347     if (no)
 348     {
 349         yaz_log(YLOG_DEBUG, "SNIPPET match: %s", wrbuf_cstr(w_snippet));
 350     }
 351     return no;
 352 }
 353
 354 void relevance_countwords(struct relevance *r, struct record_cluster *cluster,
 355                           const char *words, const char *rank,
 356                           const char *name)
 357 {
 358     int *w = r->term_frequency_vec_tmp;
 359     const char *norm_str;
 360     int i, length = 0;
 361     double lead_decay = r->lead_decay;
 362     struct word_entry *e;
 363     WRBUF wr = cluster->relevance_explain1;
 364     int printed_about_field = 0;
 365
 366     pp2_charset_token_first(r->prt, words, 0);
 367     for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
 368     {
 369         w[i] = 0;
 370         r->term_pos[i] = 0;
 371     }
 372
 373     assert(rank);
 374     while ((norm_str = pp2_charset_token_next(r->prt)))
 375     {
 376         int local_weight = 0;
 377         e = word_entry_match(r, norm_str, rank, &local_weight);
 378         if (e)
 379         {
 380             int res = e->termno;
 381             int j;
 382
 383             if (!printed_about_field)
 384             {
 385                 printed_about_field = 1;
 386                 wrbuf_printf(wr, "field=%s content=", name);
 387                 if (strlen(words) > 50)
 388                 {
 389                     wrbuf_xmlputs_n(wr, words, 49);
 390                     wrbuf_puts(wr, " ...");
 391                 }
 392                 else
 393                     wrbuf_xmlputs(wr, words);
 394                 wrbuf_puts(wr, ";\n");
 395             }
 396             assert(res < r->vec_len);
 397             w[res] += local_weight / (1 + log2(1 + lead_decay * length));
 398             wrbuf_printf(wr, "%s: w[%d] += w(%d) / "
 399                          "(1+log2(1+lead_decay(%f) * length(%d)));\n",
 400                          e->display_str, res, local_weight, lead_decay, length);
 401             j = res - 1;
 402             if (j > 0 && r->term_pos[j])
 403             {
 404                 int d = length + 1 - r->term_pos[j];
 405                 wrbuf_printf(wr, "%s: w[%d] += w[%d](%d) * follow(%f) / "
 406                              "(1+log2(d(%d));\n",
 407                              e->display_str, res, res, w[res],
 408                              r->follow_factor, d);
 409                 w[res] += w[res] * r->follow_factor / (1 + log2(d));
 410             }
 411             for (j = 0; j < r->vec_len; j++)
 412                 r->term_pos[j] = j < res ? 0 : length + 1;
 413         }
 414         length++;
 415     }
 416
 417     for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
 418     {
 419         if (length == 0 || w[i] == 0)
 420             continue;
 421         wrbuf_printf(wr, "%s: tf[%d] += w[%d](%d)", e->display_str, i, i, w[i]);
 422         switch (r->length_divide)
 423         {
 424         case 0:
 425             cluster->term_frequency_vecf[i] += (double) w[i];
 426             break;
 427         case 1:
 428             wrbuf_printf(wr, " / log2(1+length(%d))", length);
 429             cluster->term_frequency_vecf[i] +=
 430                 (double) w[i] / log2(1 + length);
 431             break;
 432         case 2:
 433             wrbuf_printf(wr, " / length(%d)", length);
 434             cluster->term_frequency_vecf[i] += (double) w[i] / length;
 435         }
 436         cluster->term_frequency_vec[i] += w[i];
 437         wrbuf_printf(wr, " (%f);\n", cluster->term_frequency_vecf[i]);
 438     }
 439
 440     cluster->term_frequency_vec[0] += length;
 441 }
 442
 443 static void pull_terms(struct relevance *res, struct ccl_rpn_node *n)
 444 {
 445     char **words;
 446     int numwords;
 447     char *ccl_field;
 448     int i;
 449
 450     switch (n->kind)
 451     {
 452     case CCL_RPN_AND:
 453     case CCL_RPN_OR:
 454     case CCL_RPN_NOT:
 455     case CCL_RPN_PROX:
 456         pull_terms(res, n->u.p[0]);
 457         pull_terms(res, n->u.p[1]);
 458         break;
 459     case CCL_RPN_TERM:
 460         nmem_strsplit(res->nmem, " ", n->u.t.term, &words, &numwords);
 461         for (i = 0; i < numwords; i++)
 462         {
 463             const char *norm_str;
 464
 465             ccl_field = nmem_strdup_null(res->nmem, n->u.t.qual);
 466
 467             pp2_charset_token_first(res->prt, words[i], 0);
 468             while ((norm_str = pp2_charset_token_next(res->prt)))
 469             {
 470                 struct word_entry **e = &res->entries;
 471                 while (*e)
 472                     e = &(*e)->next;
 473                 *e = nmem_malloc(res->nmem, sizeof(**e));
 474                 (*e)->norm_str = nmem_strdup(res->nmem, norm_str);
 475                 (*e)->ccl_field = ccl_field;
 476                 (*e)->termno = res->vec_len++;
 477                 (*e)->display_str = nmem_strdup(res->nmem, words[i]);
 478                 (*e)->next = 0;
 479             }
 480         }
 481         break;
 482     default:
 483         break;
 484     }
 485 }
 486 void relevance_clear(struct relevance *r)
 487 {
 488     if (r)
 489     {
 490         int i;
 491         for (i = 0; i < r->vec_len; i++)
 492             r->doc_frequency_vec[i] = 0;
 493     }
 494 }
 495
 496 struct relevance *relevance_create_ccl(pp2_charset_fact_t pft,
 497                                        struct ccl_rpn_node *query,
 498                                        int rank_cluster,
 499                                        double follow_factor, double lead_decay,
 500                                        int length_divide)
 501 {
 502     NMEM nmem = nmem_create();
 503     struct relevance *res = nmem_malloc(nmem, sizeof(*res));
 504
 505     res->nmem = nmem;
 506     res->entries = 0;
 507     res->vec_len = 1;
 508     res->rank_cluster = rank_cluster;
 509     res->follow_factor = follow_factor;
 510     res->lead_decay = lead_decay;
 511     res->length_divide = length_divide;
 512     res->norm = 0;
 513     res->prt = pp2_charset_token_create(pft, "relevance");
 514
 515     pull_terms(res, query);
 516
 517     res->doc_frequency_vec = nmem_malloc(nmem, res->vec_len * sizeof(int));
 518
 519     // worker array
 520     res->term_frequency_vec_tmp =
 521         nmem_malloc(res->nmem,
 522                     res->vec_len * sizeof(*res->term_frequency_vec_tmp));
 523
 524     res->term_pos =
 525         nmem_malloc(res->nmem, res->vec_len * sizeof(*res->term_pos));
 526
 527     relevance_clear(res);
 528     return res;
 529 }
 530
 531 void relevance_destroy(struct relevance **rp)
 532 {
 533     if (*rp)
 534     {
 535         pp2_charset_token_destroy((*rp)->prt);
 536         nmem_destroy((*rp)->nmem);
 537         *rp = 0;
 538     }
 539 }
 540
 541 void relevance_mergerec(struct relevance *r, struct record_cluster *dst,
 542                         const struct record_cluster *src)
 543 {
 544     int i;
 545
 546     for (i = 0; i < r->vec_len; i++)
 547         dst->term_frequency_vec[i] += src->term_frequency_vec[i];
 548
 549     for (i = 0; i < r->vec_len; i++)
 550         dst->term_frequency_vecf[i] += src->term_frequency_vecf[i];
 551 }
 552
 553 void relevance_newrec(struct relevance *r, struct record_cluster *rec)
 554 {
 555     int i;
 556
 557     // term frequency [1,..] . [0] is total length of all fields
 558     rec->term_frequency_vec =
 559         nmem_malloc(r->nmem,
 560                     r->vec_len * sizeof(*rec->term_frequency_vec));
 561     for (i = 0; i < r->vec_len; i++)
 562         rec->term_frequency_vec[i] = 0;
 563
 564     // term frequency divided by length of field [1,...]
 565     rec->term_frequency_vecf =
 566         nmem_malloc(r->nmem,
 567                     r->vec_len * sizeof(*rec->term_frequency_vecf));
 568     for (i = 0; i < r->vec_len; i++)
 569         rec->term_frequency_vecf[i] = 0.0;
 570 }
 571
 572 void relevance_donerecord(struct relevance *r, struct record_cluster *cluster)
 573 {
 574     int i;
 575
 576     for (i = 1; i < r->vec_len; i++)
 577         if (cluster->term_frequency_vec[i] > 0)
 578             r->doc_frequency_vec[i]++;
 579
 580     r->doc_frequency_vec[0]++;
 581 }
 582
 583
 584
 585 // Prepare for a relevance-sorted read
 586 void relevance_prepare_read(struct relevance *rel, struct reclist *reclist)
 587 {
 588     int i;
 589     float *idfvec = xmalloc(rel->vec_len * sizeof(float));
 590
 591     reclist_enter(reclist);
 592
 593     // Calculate document frequency vector for each term.
 594     for (i = 1; i < rel->vec_len; i++)
 595     {
 596         if (!rel->doc_frequency_vec[i])
 597             idfvec[i] = 0;
 598         else
 599         {
 600             /* add one to nominator idf(t,D) to ensure a value > 0 */
 601             idfvec[i] = log((float) (1 + rel->doc_frequency_vec[0]) /
 602                             rel->doc_frequency_vec[i]);
 603         }
 604     }
 605     // Calculate relevance for each document
 606     while (1)
 607     {
 608         int relevance = 0;
 609         WRBUF w;
 610         struct word_entry *e = rel->entries;
 611         struct record_cluster *rec = reclist_read_record(reclist);
 612         if (!rec)
 613             break;
 614         w = rec->relevance_explain2;
 615         wrbuf_rewind(w);
 616         wrbuf_puts(w, "relevance = 0;\n");
 617         for (i = 1; i < rel->vec_len; i++)
 618         {
 619             float termfreq = (float) rec->term_frequency_vecf[i];
 620             int add = 100000 * termfreq * idfvec[i];
 621
 622             wrbuf_printf(w, "idf[%d] = log(((1 + total(%d))/termoccur(%d));\n",
 623                          i, rel->doc_frequency_vec[0],
 624                          rel->doc_frequency_vec[i]);
 625             wrbuf_printf(w, "%s: relevance += 100000 * tf[%d](%f) * "
 626                          "idf[%d](%f) (%d);\n",
 627                          e->display_str, i, termfreq, i, idfvec[i], add);
 628             relevance += add;
 629             e = e->next;
 630         }
 631         if (!rel->rank_cluster)
 632         {
 633             struct record *record;
 634             int cluster_size = 0;
 635
 636             for (record = rec->records; record; record = record->next)
 637                 cluster_size++;
 638
 639             wrbuf_printf(w, "score = relevance(%d)/cluster_size(%d);\n",
 640                          relevance, cluster_size);
 641             relevance /= cluster_size;
 642         }
 643         else
 644         {
 645             wrbuf_printf(w, "score = relevance(%d);\n", relevance);
 646         }
 647         rec->relevance_score = relevance;
 648
 649         // Build the normalizing structures
 650         // List of (sub)records for each target
 651         setup_norm_record( rel, rec );
 652
 653         // TODO - Loop again, merge individual record scores into clusters
 654         // Can I reset the reclist, or can I leave and enter without race conditions?
 655
 656     } // cluster loop
 657
 658     normalize_scores(rel);
 659
 660     reclist_leave(reclist);
 661     xfree(idfvec);
 662
 663 }
 664
 665 /*
 666  * Local variables:
 667  * c-basic-offset: 4
 668  * c-file-style: "Stroustrup"
 669  * indent-tabs-mode: nil
 670  * End:
 671  * vim: shiftwidth=4 tabstop=8 expandtab
 672  */
 673