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             {
 106                 break;
 107                 if (!highlight)
 108                 {
 109                     highlight = 1;
 110                     wrbuf_puts(w_snippet, "<match>");
 111                 }
 112                 break;
 113             }
 114         }
 115         if (entries)
 116         {
 117             if (!highlight)
 118             {
 119                 highlight = 1;
 120                 wrbuf_puts(w_snippet, "<match>");
 121                 no++;
 122             }
 123         }
 124         else
 125         {
 126             if (highlight)
 127             {
 128                 highlight = 0;
 129                 wrbuf_puts(w_snippet, "</match>");
 130             }
 131         }
 132         wrbuf_xmlputs_n(w_snippet, words + org_start, org_len);
 133     }
 134     if (highlight)
 135         wrbuf_puts(w_snippet, "</match>");
 136     if (no)
 137     {
 138         yaz_log(YLOG_DEBUG, "SNIPPET match: %s", wrbuf_cstr(w_snippet));
 139     }
 140     return no;
 141 }
 142
 143 void relevance_countwords(struct relevance *r, struct record_cluster *cluster,
 144                           const char *words, const char *rank,
 145                           const char *name)
 146 {
 147     int *w = r->term_frequency_vec_tmp;
 148     const char *norm_str;
 149     int i, length = 0;
 150     double lead_decay = r->lead_decay;
 151     struct word_entry *e;
 152     WRBUF wr = cluster->relevance_explain1;
 153     int printed_about_field = 0;
 154
 155     pp2_charset_token_first(r->prt, words, 0);
 156     for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
 157     {
 158         w[i] = 0;
 159         r->term_pos[i] = 0;
 160     }
 161
 162     assert(rank);
 163     while ((norm_str = pp2_charset_token_next(r->prt)))
 164     {
 165         int local_weight = 0;
 166         e = word_entry_match(r, norm_str, rank, &local_weight);
 167         if (e)
 168         {
 169             int res = e->termno;
 170             int j;
 171
 172             if (!printed_about_field)
 173             {
 174                 printed_about_field = 1;
 175                 wrbuf_printf(wr, "field=%s content=", name);
 176                 if (strlen(words) > 50)
 177                 {
 178                     wrbuf_xmlputs_n(wr, words, 49);
 179                     wrbuf_puts(wr, " ...");
 180                 }
 181                 else
 182                     wrbuf_xmlputs(wr, words);
 183                 wrbuf_puts(wr, ";\n");
 184             }
 185             assert(res < r->vec_len);
 186             w[res] += local_weight / (1 + log2(1 + lead_decay * length));
 187             wrbuf_printf(wr, "%s: w[%d] += w(%d) / "
 188                          "(1+log2(1+lead_decay(%f) * length(%d)));\n",
 189                          e->display_str, res, local_weight, lead_decay, length);
 190             j = res - 1;
 191             if (j > 0 && r->term_pos[j])
 192             {
 193                 int d = length + 1 - r->term_pos[j];
 194                 wrbuf_printf(wr, "%s: w[%d] += w[%d](%d) * follow(%f) / "
 195                              "(1+log2(d(%d));\n",
 196                              e->display_str, res, res, w[res],
 197                              r->follow_factor, d);
 198                 w[res] += w[res] * r->follow_factor / (1 + log2(d));
 199             }
 200             for (j = 0; j < r->vec_len; j++)
 201                 r->term_pos[j] = j < res ? 0 : length + 1;
 202         }
 203         length++;
 204     }
 205
 206     for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
 207     {
 208         if (length == 0 || w[i] == 0)
 209             continue;
 210         wrbuf_printf(wr, "%s: tf[%d] += w[%d](%d)", e->display_str, i, i, w[i]);
 211         switch (r->length_divide)
 212         {
 213         case 0:
 214             cluster->term_frequency_vecf[i] += (double) w[i];
 215             break;
 216         case 1:
 217             wrbuf_printf(wr, " / log2(1+length(%d))", length);
 218             cluster->term_frequency_vecf[i] +=
 219                 (double) w[i] / log2(1 + length);
 220             break;
 221         case 2:
 222             wrbuf_printf(wr, " / length(%d)", length);
 223             cluster->term_frequency_vecf[i] += (double) w[i] / length;
 224         }
 225         cluster->term_frequency_vec[i] += w[i];
 226         wrbuf_printf(wr, " (%f);\n", cluster->term_frequency_vecf[i]);
 227     }
 228
 229     cluster->term_frequency_vec[0] += length;
 230 }
 231
 232 static void pull_terms(struct relevance *res, struct ccl_rpn_node *n)
 233 {
 234     char **words;
 235     int numwords;
 236     char *ccl_field;
 237     int i;
 238
 239     switch (n->kind)
 240     {
 241     case CCL_RPN_AND:
 242     case CCL_RPN_OR:
 243     case CCL_RPN_NOT:
 244     case CCL_RPN_PROX:
 245         pull_terms(res, n->u.p[0]);
 246         pull_terms(res, n->u.p[1]);
 247         break;
 248     case CCL_RPN_TERM:
 249         nmem_strsplit(res->nmem, " ", n->u.t.term, &words, &numwords);
 250         for (i = 0; i < numwords; i++)
 251         {
 252             const char *norm_str;
 253
 254             ccl_field = nmem_strdup_null(res->nmem, n->u.t.qual);
 255
 256             pp2_charset_token_first(res->prt, words[i], 0);
 257             while ((norm_str = pp2_charset_token_next(res->prt)))
 258             {
 259                 struct word_entry **e = &res->entries;
 260                 while (*e)
 261                     e = &(*e)->next;
 262                 *e = nmem_malloc(res->nmem, sizeof(**e));
 263                 (*e)->norm_str = nmem_strdup(res->nmem, norm_str);
 264                 (*e)->ccl_field = ccl_field;
 265                 (*e)->termno = res->vec_len++;
 266                 (*e)->display_str = nmem_strdup(res->nmem, words[i]);
 267                 (*e)->next = 0;
 268             }
 269         }
 270         break;
 271     default:
 272         break;
 273     }
 274 }
 275 void relevance_clear(struct relevance *r)
 276 {
 277     if (r)
 278     {
 279         int i;
 280         for (i = 0; i < r->vec_len; i++)
 281             r->doc_frequency_vec[i] = 0;
 282     }
 283 }
 284
 285 struct relevance *relevance_create_ccl(pp2_charset_fact_t pft,
 286                                        struct ccl_rpn_node *query,
 287                                        int rank_cluster,
 288                                        double follow_factor, double lead_decay,
 289                                        int length_divide)
 290 {
 291     NMEM nmem = nmem_create();
 292     struct relevance *res = nmem_malloc(nmem, sizeof(*res));
 293
 294     res->nmem = nmem;
 295     res->entries = 0;
 296     res->vec_len = 1;
 297     res->rank_cluster = rank_cluster;
 298     res->follow_factor = follow_factor;
 299     res->lead_decay = lead_decay;
 300     res->length_divide = length_divide;
 301     res->prt = pp2_charset_token_create(pft, "relevance");
 302
 303     pull_terms(res, query);
 304
 305     res->doc_frequency_vec = nmem_malloc(nmem, res->vec_len * sizeof(int));
 306
 307     // worker array
 308     res->term_frequency_vec_tmp =
 309         nmem_malloc(res->nmem,
 310                     res->vec_len * sizeof(*res->term_frequency_vec_tmp));
 311
 312     res->term_pos =
 313         nmem_malloc(res->nmem, res->vec_len * sizeof(*res->term_pos));
 314
 315     relevance_clear(res);
 316     return res;
 317 }
 318
 319 void relevance_destroy(struct relevance **rp)
 320 {
 321     if (*rp)
 322     {
 323         pp2_charset_token_destroy((*rp)->prt);
 324         nmem_destroy((*rp)->nmem);
 325         *rp = 0;
 326     }
 327 }
 328
 329 void relevance_newrec(struct relevance *r, struct record_cluster *rec)
 330 {
 331     if (!rec->term_frequency_vec)
 332     {
 333         int i;
 334
 335         // term frequency [1,..] . [0] is total length of all fields
 336         rec->term_frequency_vec =
 337             nmem_malloc(r->nmem,
 338                         r->vec_len * sizeof(*rec->term_frequency_vec));
 339         for (i = 0; i < r->vec_len; i++)
 340             rec->term_frequency_vec[i] = 0;
 341
 342         // term frequency divided by length of field [1,...]
 343         rec->term_frequency_vecf =
 344             nmem_malloc(r->nmem,
 345                         r->vec_len * sizeof(*rec->term_frequency_vecf));
 346         for (i = 0; i < r->vec_len; i++)
 347             rec->term_frequency_vecf[i] = 0.0;
 348     }
 349 }
 350
 351 void relevance_donerecord(struct relevance *r, struct record_cluster *cluster)
 352 {
 353     int i;
 354
 355     for (i = 1; i < r->vec_len; i++)
 356         if (cluster->term_frequency_vec[i] > 0)
 357             r->doc_frequency_vec[i]++;
 358
 359     r->doc_frequency_vec[0]++;
 360 }
 361
 362 // Prepare for a relevance-sorted read
 363 void relevance_prepare_read(struct relevance *rel, struct reclist *reclist)
 364 {
 365     int i;
 366     float *idfvec = xmalloc(rel->vec_len * sizeof(float));
 367
 368     reclist_enter(reclist);
 369     // Calculate document frequency vector for each term.
 370     for (i = 1; i < rel->vec_len; i++)
 371     {
 372         if (!rel->doc_frequency_vec[i])
 373             idfvec[i] = 0;
 374         else
 375         {
 376             /* add one to nominator idf(t,D) to ensure a value > 0 */
 377             idfvec[i] = log((float) (1 + rel->doc_frequency_vec[0]) /
 378                             rel->doc_frequency_vec[i]);
 379         }
 380     }
 381     // Calculate relevance for each document
 382     while (1)
 383     {
 384         int relevance = 0;
 385         WRBUF w;
 386         struct word_entry *e = rel->entries;
 387         struct record_cluster *rec = reclist_read_record(reclist);
 388         if (!rec)
 389             break;
 390         w = rec->relevance_explain2;
 391         wrbuf_rewind(w);
 392         wrbuf_puts(w, "relevance = 0;\n");
 393         for (i = 1; i < rel->vec_len; i++)
 394         {
 395             float termfreq = (float) rec->term_frequency_vecf[i];
 396             int add = 100000 * termfreq * idfvec[i];
 397
 398             wrbuf_printf(w, "idf[%d] = log(((1 + total(%d))/termoccur(%d));\n",
 399                          i, rel->doc_frequency_vec[0],
 400                          rel->doc_frequency_vec[i]);
 401             wrbuf_printf(w, "%s: relevance += 100000 * tf[%d](%f) * "
 402                          "idf[%d](%f) (%d);\n",
 403                          e->display_str, i, termfreq, i, idfvec[i], add);
 404             relevance += add;
 405             e = e->next;
 406         }
 407         if (!rel->rank_cluster)
 408         {
 409             struct record *record;
 410             int cluster_size = 0;
 411
 412             for (record = rec->records; record; record = record->next)
 413                 cluster_size++;
 414
 415             wrbuf_printf(w, "score = relevance(%d)/cluster_size(%d);\n",
 416                          relevance, cluster_size);
 417             relevance /= cluster_size;
 418         }
 419         else
 420         {
 421             wrbuf_printf(w, "score = relevance(%d);\n", relevance);
 422         }
 423         rec->relevance_score = relevance;
 424     }
 425     reclist_leave(reclist);
 426     xfree(idfvec);
 427 }
 428
 429 /*
 430  * Local variables:
 431  * c-basic-offset: 4
 432  * c-file-style: "Stroustrup"
 433  * indent-tabs-mode: nil
 434  * End:
 435  * vim: shiftwidth=4 tabstop=8 expandtab
 436  */
 437