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 "pazpar2_config.h"
  29 #include "relevance.h"
  30 #include "session.h"
  31
  32 #ifdef WIN32
  33 #define log2(x) (log(x)/log(2))
  34 #endif
  35
  36 struct relevance
  37 {
  38     int *doc_frequency_vec;
  39     int *term_frequency_vec_tmp;
  40     int *term_pos;
  41     int vec_len;
  42     struct word_entry *entries;
  43     pp2_charset_token_t prt;
  44     int rank_cluster;
  45     double follow_factor;
  46     double lead_decay;
  47     int length_divide;
  48     NMEM nmem;
  49 };
  50
  51 struct word_entry {
  52     const char *norm_str;
  53     const char *display_str;
  54     int termno;
  55     char *ccl_field;
  56     struct word_entry *next;
  57 };
  58
  59 static struct word_entry *word_entry_match(struct relevance *r,
  60                                            const char *norm_str,
  61                                            const char *rank, int *weight)
  62 {
  63     int i = 1;
  64     struct word_entry *entries = r->entries;
  65     for (; entries; entries = entries->next, i++)
  66     {
  67         if (*norm_str && !strcmp(norm_str, entries->norm_str))
  68         {
  69             const char *cp = 0;
  70             int no_read = 0;
  71             sscanf(rank, "%d%n", weight, &no_read);
  72             rank += no_read;
  73             while (*rank == ' ')
  74                 rank++;
  75             if (no_read > 0 && (cp = strchr(rank, ' ')))
  76             {
  77                 if ((cp - rank) == strlen(entries->ccl_field) &&
  78                     memcmp(entries->ccl_field, rank, cp - rank) == 0)
  79                     *weight = atoi(cp + 1);
  80             }
  81             return entries;
  82         }
  83     }
  84     return 0;
  85 }
  86
  87 int relevance_snippet(struct relevance *r,
  88                       const char *words, const char *name,
  89                       WRBUF w_snippet)
  90 {
  91     int no = 0;
  92     const char *norm_str;
  93     int highlight = 0;
  94
  95     pp2_charset_token_first(r->prt, words, 0);
  96     while ((norm_str = pp2_charset_token_next(r->prt)))
  97     {
  98         size_t org_start, org_len;
  99         struct word_entry *entries = r->entries;
 100         int i;
 101
 102         pp2_get_org(r->prt, &org_start, &org_len);
 103         for (; entries; entries = entries->next, i++)
 104         {
 105             if (*norm_str && !strcmp(norm_str, entries->norm_str))
 106                 break;
 107         }
 108         if (entries)
 109         {
 110             if (!highlight)
 111             {
 112                 highlight = 1;
 113                 wrbuf_puts(w_snippet, "<match>");
 114                 no++;
 115             }
 116         }
 117         else
 118         {
 119             if (highlight)
 120             {
 121                 highlight = 0;
 122                 wrbuf_puts(w_snippet, "</match>");
 123             }
 124         }
 125         wrbuf_xmlputs_n(w_snippet, words + org_start, org_len);
 126     }
 127     if (highlight)
 128         wrbuf_puts(w_snippet, "</match>");
 129     if (no)
 130     {
 131         yaz_log(YLOG_DEBUG, "SNIPPET match: %s", wrbuf_cstr(w_snippet));
 132     }
 133     return no;
 134 }
 135
 136 void relevance_countwords(struct relevance *r, struct record_cluster *cluster,
 137                           const char *words, const char *rank,
 138                           const char *name)
 139 {
 140     int *w = r->term_frequency_vec_tmp;
 141     const char *norm_str;
 142     int i, length = 0;
 143     double lead_decay = r->lead_decay;
 144     struct word_entry *e;
 145     WRBUF wr = cluster->relevance_explain1;
 146     int printed_about_field = 0;
 147
 148     pp2_charset_token_first(r->prt, words, 0);
 149     for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
 150     {
 151         w[i] = 0;
 152         r->term_pos[i] = 0;
 153     }
 154
 155     assert(rank);
 156     while ((norm_str = pp2_charset_token_next(r->prt)))
 157     {
 158         int local_weight = 0;
 159         e = word_entry_match(r, norm_str, rank, &local_weight);
 160         if (e)
 161         {
 162             int res = e->termno;
 163             int j;
 164
 165             if (!printed_about_field)
 166             {
 167                 printed_about_field = 1;
 168                 wrbuf_printf(wr, "field=%s content=", name);
 169                 if (strlen(words) > 50)
 170                 {
 171                     wrbuf_xmlputs_n(wr, words, 49);
 172                     wrbuf_puts(wr, " ...");
 173                 }
 174                 else
 175                     wrbuf_xmlputs(wr, words);
 176                 wrbuf_puts(wr, ";\n");
 177             }
 178             assert(res < r->vec_len);
 179             w[res] += local_weight / (1 + log2(1 + lead_decay * length));
 180             wrbuf_printf(wr, "%s: w[%d] += w(%d) / "
 181                          "(1+log2(1+lead_decay(%f) * length(%d)));\n",
 182                          e->display_str, res, local_weight, lead_decay, length);
 183             j = res - 1;
 184             if (j > 0 && r->term_pos[j])
 185             {
 186                 int d = length + 1 - r->term_pos[j];
 187                 wrbuf_printf(wr, "%s: w[%d] += w[%d](%d) * follow(%f) / "
 188                              "(1+log2(d(%d));\n",
 189                              e->display_str, res, res, w[res],
 190                              r->follow_factor, d);
 191                 w[res] += w[res] * r->follow_factor / (1 + log2(d));
 192             }
 193             for (j = 0; j < r->vec_len; j++)
 194                 r->term_pos[j] = j < res ? 0 : length + 1;
 195         }
 196         length++;
 197     }
 198
 199     for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
 200     {
 201         if (length == 0 || w[i] == 0)
 202             continue;
 203         wrbuf_printf(wr, "%s: tf[%d] += w[%d](%d)", e->display_str, i, i, w[i]);
 204         switch (r->length_divide)
 205         {
 206         case 0:
 207             cluster->term_frequency_vecf[i] += (double) w[i];
 208             break;
 209         case 1:
 210             wrbuf_printf(wr, " / log2(1+length(%d))", length);
 211             cluster->term_frequency_vecf[i] +=
 212                 (double) w[i] / log2(1 + length);
 213             break;
 214         case 2:
 215             wrbuf_printf(wr, " / length(%d)", length);
 216             cluster->term_frequency_vecf[i] += (double) w[i] / length;
 217         }
 218         cluster->term_frequency_vec[i] += w[i];
 219         wrbuf_printf(wr, " (%f);\n", cluster->term_frequency_vecf[i]);
 220     }
 221
 222     cluster->term_frequency_vec[0] += length;
 223 }
 224
 225 static void pull_terms(struct relevance *res, struct ccl_rpn_node *n)
 226 {
 227     char **words;
 228     int numwords;
 229     char *ccl_field;
 230     int i;
 231
 232     switch (n->kind)
 233     {
 234     case CCL_RPN_AND:
 235     case CCL_RPN_OR:
 236     case CCL_RPN_NOT:
 237     case CCL_RPN_PROX:
 238         pull_terms(res, n->u.p[0]);
 239         pull_terms(res, n->u.p[1]);
 240         break;
 241     case CCL_RPN_TERM:
 242         nmem_strsplit(res->nmem, " ", n->u.t.term, &words, &numwords);
 243         for (i = 0; i < numwords; i++)
 244         {
 245             const char *norm_str;
 246
 247             ccl_field = nmem_strdup_null(res->nmem, n->u.t.qual);
 248
 249             pp2_charset_token_first(res->prt, words[i], 0);
 250             while ((norm_str = pp2_charset_token_next(res->prt)))
 251             {
 252                 struct word_entry **e = &res->entries;
 253                 while (*e)
 254                     e = &(*e)->next;
 255                 *e = nmem_malloc(res->nmem, sizeof(**e));
 256                 (*e)->norm_str = nmem_strdup(res->nmem, norm_str);
 257                 (*e)->ccl_field = ccl_field;
 258                 (*e)->termno = res->vec_len++;
 259                 (*e)->display_str = nmem_strdup(res->nmem, words[i]);
 260                 (*e)->next = 0;
 261             }
 262         }
 263         break;
 264     default:
 265         break;
 266     }
 267 }
 268 void relevance_clear(struct relevance *r)
 269 {
 270     if (r)
 271     {
 272         int i;
 273         for (i = 0; i < r->vec_len; i++)
 274             r->doc_frequency_vec[i] = 0;
 275     }
 276 }
 277
 278 struct relevance *relevance_create_ccl(pp2_charset_fact_t pft,
 279                                        struct ccl_rpn_node *query,
 280                                        int rank_cluster,
 281                                        double follow_factor, double lead_decay,
 282                                        int length_divide)
 283 {
 284     NMEM nmem = nmem_create();
 285     struct relevance *res = nmem_malloc(nmem, sizeof(*res));
 286
 287     res->nmem = nmem;
 288     res->entries = 0;
 289     res->vec_len = 1;
 290     res->rank_cluster = rank_cluster;
 291     res->follow_factor = follow_factor;
 292     res->lead_decay = lead_decay;
 293     res->length_divide = length_divide;
 294     res->prt = pp2_charset_token_create(pft, "relevance");
 295
 296     pull_terms(res, query);
 297
 298     res->doc_frequency_vec = nmem_malloc(nmem, res->vec_len * sizeof(int));
 299
 300     // worker array
 301     res->term_frequency_vec_tmp =
 302         nmem_malloc(res->nmem,
 303                     res->vec_len * sizeof(*res->term_frequency_vec_tmp));
 304
 305     res->term_pos =
 306         nmem_malloc(res->nmem, res->vec_len * sizeof(*res->term_pos));
 307
 308     relevance_clear(res);
 309     return res;
 310 }
 311
 312 void relevance_destroy(struct relevance **rp)
 313 {
 314     if (*rp)
 315     {
 316         pp2_charset_token_destroy((*rp)->prt);
 317         nmem_destroy((*rp)->nmem);
 318         *rp = 0;
 319     }
 320 }
 321
 322 void relevance_newrec(struct relevance *r, struct record_cluster *rec)
 323 {
 324     if (!rec->term_frequency_vec)
 325     {
 326         int i;
 327
 328         // term frequency [1,..] . [0] is total length of all fields
 329         rec->term_frequency_vec =
 330             nmem_malloc(r->nmem,
 331                         r->vec_len * sizeof(*rec->term_frequency_vec));
 332         for (i = 0; i < r->vec_len; i++)
 333             rec->term_frequency_vec[i] = 0;
 334
 335         // term frequency divided by length of field [1,...]
 336         rec->term_frequency_vecf =
 337             nmem_malloc(r->nmem,
 338                         r->vec_len * sizeof(*rec->term_frequency_vecf));
 339         for (i = 0; i < r->vec_len; i++)
 340             rec->term_frequency_vecf[i] = 0.0;
 341     }
 342 }
 343
 344 void relevance_donerecord(struct relevance *r, struct record_cluster *cluster)
 345 {
 346     int i;
 347
 348     for (i = 1; i < r->vec_len; i++)
 349         if (cluster->term_frequency_vec[i] > 0)
 350             r->doc_frequency_vec[i]++;
 351
 352     r->doc_frequency_vec[0]++;
 353 }
 354
 355 // Prepare for a relevance-sorted read
 356 void relevance_prepare_read(struct relevance *rel, struct reclist *reclist,
 357                             enum conf_sortkey_type type)
 358 {
 359     int i;
 360     float *idfvec = xmalloc(rel->vec_len * sizeof(float));
 361
 362     reclist_enter(reclist);
 363     // Calculate document frequency vector for each term.
 364     for (i = 1; i < rel->vec_len; i++)
 365     {
 366         if (!rel->doc_frequency_vec[i])
 367             idfvec[i] = 0;
 368         else
 369         {
 370             /* add one to nominator idf(t,D) to ensure a value > 0 */
 371             idfvec[i] = log((float) (1 + rel->doc_frequency_vec[0]) /
 372                             rel->doc_frequency_vec[i]);
 373         }
 374     }
 375     // Calculate relevance for each document
 376     while (1)
 377     {
 378         int relevance = 0;
 379         WRBUF w;
 380         struct word_entry *e = rel->entries;
 381         struct record_cluster *rec = reclist_read_record(reclist);
 382         if (!rec)
 383             break;
 384         w = rec->relevance_explain2;
 385         wrbuf_rewind(w);
 386         wrbuf_puts(w, "relevance = 0;\n");
 387         for (i = 1; i < rel->vec_len; i++)
 388         {
 389             float termfreq = (float) rec->term_frequency_vecf[i];
 390             int add = 100000 * termfreq * idfvec[i];
 391
 392             wrbuf_printf(w, "idf[%d] = log(((1 + total(%d))/termoccur(%d));\n",
 393                          i, rel->doc_frequency_vec[0],
 394                          rel->doc_frequency_vec[i]);
 395             wrbuf_printf(w, "%s: relevance += 100000 * tf[%d](%f) * "
 396                          "idf[%d](%f) (%d);\n",
 397                          e->display_str, i, termfreq, i, idfvec[i], add);
 398             relevance += add;
 399             e = e->next;
 400         }
 401         if (!rel->rank_cluster)
 402         {
 403             struct record *record;
 404             int cluster_size = 0;
 405
 406             for (record = rec->records; record; record = record->next)
 407                 cluster_size++;
 408
 409             wrbuf_printf(w, "score = relevance(%d)/cluster_size(%d);\n",
 410                          relevance, cluster_size);
 411             relevance /= cluster_size;
 412         }
 413         else
 414         {
 415             wrbuf_printf(w, "score = relevance(%d);\n", relevance);
 416         }
 417         if (0 && type == Metadata_sortkey_relevance_h)
 418             relevance *= 2;
 419         rec->relevance_score = relevance;
 420     }
 421     reclist_leave(reclist);
 422     xfree(idfvec);
 423 }
 424
 425 /*
 426  * Local variables:
 427  * c-basic-offset: 4
 428  * c-file-style: "Stroustrup"
 429  * indent-tabs-mode: nil
 430  * End:
 431  * vim: shiftwidth=4 tabstop=8 expandtab
 432  */
 433