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