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 void relevance_countwords(struct relevance *r, struct record_cluster *cluster,
  87                           const char *words, const char *rank,
  88                           const char *name)
  89 {
  90     int *w = r->term_frequency_vec_tmp;
  91     const char *norm_str;
  92     int i, length = 0;
  93     double lead_decay = r->lead_decay;
  94     struct word_entry *e;
  95     WRBUF wr = cluster->relevance_explain1;
  96     int printed_about_field = 0;
  97
  98     pp2_charset_token_first(r->prt, words, 0);
  99     for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
 100     {
 101         w[i] = 0;
 102         r->term_pos[i] = 0;
 103     }
 104
 105     assert(rank);
 106     while ((norm_str = pp2_charset_token_next(r->prt)))
 107     {
 108         int local_weight = 0;
 109         e = word_entry_match(r, norm_str, rank, &local_weight);
 110         if (e)
 111         {
 112             int res = e->termno;
 113             int j;
 114
 115             if (!printed_about_field)
 116             {
 117                 printed_about_field = 1;
 118                 wrbuf_printf(wr, "field=%s content=", name);
 119                 if (strlen(words) > 50)
 120                 {
 121                     wrbuf_xmlputs_n(wr, words, 49);
 122                     wrbuf_puts(wr, " ...");
 123                 }
 124                 else
 125                     wrbuf_xmlputs(wr, words);
 126                 wrbuf_puts(wr, ";\n");
 127             }
 128             assert(res < r->vec_len);
 129             w[res] += local_weight / (1 + log2(1 + lead_decay * length));
 130             wrbuf_printf(wr, "%s: w[%d] += w(%d) / "
 131                          "(1+log2(1+lead_decay(%f) * length(%d)));\n",
 132                          e->display_str, res, local_weight, lead_decay, length);
 133             j = res - 1;
 134             if (j > 0 && r->term_pos[j])
 135             {
 136                 int d = length + 1 - r->term_pos[j];
 137                 wrbuf_printf(wr, "%s: w[%d] += w[%d](%d) * follow(%f) / "
 138                              "(1+log2(d(%d));\n",
 139                              e->display_str, res, res, w[res],
 140                              r->follow_factor, d);
 141                 w[res] += w[res] * r->follow_factor / (1 + log2(d));
 142             }
 143             for (j = 0; j < r->vec_len; j++)
 144                 r->term_pos[j] = j < res ? 0 : length + 1;
 145         }
 146         length++;
 147     }
 148
 149     for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
 150     {
 151         if (length == 0 || w[i] == 0)
 152             continue;
 153         wrbuf_printf(wr, "%s: tf[%d] += w[%d](%d)", e->display_str, i, i, w[i]);
 154         switch (r->length_divide)
 155         {
 156         case 0:
 157             cluster->term_frequency_vecf[i] += (double) w[i];
 158             break;
 159         case 1:
 160             wrbuf_printf(wr, " / log2(1+length(%d))", length);
 161             cluster->term_frequency_vecf[i] +=
 162                 (double) w[i] / log2(1 + length);
 163             break;
 164         case 2:
 165             wrbuf_printf(wr, " / length(%d)", length);
 166             cluster->term_frequency_vecf[i] += (double) w[i] / length;
 167         }
 168         cluster->term_frequency_vec[i] += w[i];
 169         wrbuf_printf(wr, " (%f);\n", cluster->term_frequency_vecf[i]);
 170     }
 171
 172     cluster->term_frequency_vec[0] += length;
 173 }
 174
 175 static void pull_terms(struct relevance *res, struct ccl_rpn_node *n)
 176 {
 177     char **words;
 178     int numwords;
 179     char *ccl_field;
 180     int i;
 181
 182     switch (n->kind)
 183     {
 184     case CCL_RPN_AND:
 185     case CCL_RPN_OR:
 186     case CCL_RPN_NOT:
 187     case CCL_RPN_PROX:
 188         pull_terms(res, n->u.p[0]);
 189         pull_terms(res, n->u.p[1]);
 190         break;
 191     case CCL_RPN_TERM:
 192         nmem_strsplit(res->nmem, " ", n->u.t.term, &words, &numwords);
 193         for (i = 0; i < numwords; i++)
 194         {
 195             const char *norm_str;
 196
 197             ccl_field = nmem_strdup_null(res->nmem, n->u.t.qual);
 198
 199             pp2_charset_token_first(res->prt, words[i], 0);
 200             while ((norm_str = pp2_charset_token_next(res->prt)))
 201             {
 202                 struct word_entry **e = &res->entries;
 203                 while (*e)
 204                     e = &(*e)->next;
 205                 *e = nmem_malloc(res->nmem, sizeof(**e));
 206                 (*e)->norm_str = nmem_strdup(res->nmem, norm_str);
 207                 (*e)->ccl_field = ccl_field;
 208                 (*e)->termno = res->vec_len++;
 209                 (*e)->display_str = nmem_strdup(res->nmem, words[i]);
 210                 (*e)->next = 0;
 211             }
 212         }
 213         break;
 214     default:
 215         break;
 216     }
 217 }
 218 void relevance_clear(struct relevance *r)
 219 {
 220     if (r)
 221     {
 222         int i;
 223         for (i = 0; i < r->vec_len; i++)
 224             r->doc_frequency_vec[i] = 0;
 225     }
 226 }
 227
 228 struct relevance *relevance_create_ccl(pp2_charset_fact_t pft,
 229                                        struct ccl_rpn_node *query,
 230                                        int rank_cluster,
 231                                        double follow_factor, double lead_decay,
 232                                        int length_divide)
 233 {
 234     NMEM nmem = nmem_create();
 235     struct relevance *res = nmem_malloc(nmem, sizeof(*res));
 236
 237     res->nmem = nmem;
 238     res->entries = 0;
 239     res->vec_len = 1;
 240     res->rank_cluster = rank_cluster;
 241     res->follow_factor = follow_factor;
 242     res->lead_decay = lead_decay;
 243     res->length_divide = length_divide;
 244     res->prt = pp2_charset_token_create(pft, "relevance");
 245
 246     pull_terms(res, query);
 247
 248     res->doc_frequency_vec = nmem_malloc(nmem, res->vec_len * sizeof(int));
 249
 250     // worker array
 251     res->term_frequency_vec_tmp =
 252         nmem_malloc(res->nmem,
 253                     res->vec_len * sizeof(*res->term_frequency_vec_tmp));
 254
 255     res->term_pos =
 256         nmem_malloc(res->nmem, res->vec_len * sizeof(*res->term_pos));
 257
 258     relevance_clear(res);
 259     return res;
 260 }
 261
 262 void relevance_destroy(struct relevance **rp)
 263 {
 264     if (*rp)
 265     {
 266         pp2_charset_token_destroy((*rp)->prt);
 267         nmem_destroy((*rp)->nmem);
 268         *rp = 0;
 269     }
 270 }
 271
 272 void relevance_newrec(struct relevance *r, struct record_cluster *rec)
 273 {
 274     if (!rec->term_frequency_vec)
 275     {
 276         int i;
 277
 278         // term frequency [1,..] . [0] is total length of all fields
 279         rec->term_frequency_vec =
 280             nmem_malloc(r->nmem,
 281                         r->vec_len * sizeof(*rec->term_frequency_vec));
 282         for (i = 0; i < r->vec_len; i++)
 283             rec->term_frequency_vec[i] = 0;
 284
 285         // term frequency divided by length of field [1,...]
 286         rec->term_frequency_vecf =
 287             nmem_malloc(r->nmem,
 288                         r->vec_len * sizeof(*rec->term_frequency_vecf));
 289         for (i = 0; i < r->vec_len; i++)
 290             rec->term_frequency_vecf[i] = 0.0;
 291     }
 292 }
 293
 294 void relevance_donerecord(struct relevance *r, struct record_cluster *cluster)
 295 {
 296     int i;
 297
 298     for (i = 1; i < r->vec_len; i++)
 299         if (cluster->term_frequency_vec[i] > 0)
 300             r->doc_frequency_vec[i]++;
 301
 302     r->doc_frequency_vec[0]++;
 303 }
 304
 305 // Prepare for a relevance-sorted read
 306 void relevance_prepare_read(struct relevance *rel, struct reclist *reclist)
 307 {
 308     int i;
 309     float *idfvec = xmalloc(rel->vec_len * sizeof(float));
 310
 311     reclist_enter(reclist);
 312     // Calculate document frequency vector for each term.
 313     for (i = 1; i < rel->vec_len; i++)
 314     {
 315         if (!rel->doc_frequency_vec[i])
 316             idfvec[i] = 0;
 317         else
 318         {
 319             /* add one to nominator idf(t,D) to ensure a value > 0 */
 320             idfvec[i] = log((float) (1 + rel->doc_frequency_vec[0]) /
 321                             rel->doc_frequency_vec[i]);
 322         }
 323     }
 324     // Calculate relevance for each document
 325     while (1)
 326     {
 327         int relevance = 0;
 328         WRBUF w;
 329         struct word_entry *e = rel->entries;
 330         struct record_cluster *rec = reclist_read_record(reclist);
 331         if (!rec)
 332             break;
 333         w = rec->relevance_explain2;
 334         wrbuf_rewind(w);
 335         wrbuf_puts(w, "relevance = 0;\n");
 336         for (i = 1; i < rel->vec_len; i++)
 337         {
 338             float termfreq = (float) rec->term_frequency_vecf[i];
 339             int add = 100000 * termfreq * idfvec[i];
 340
 341             wrbuf_printf(w, "idf[%d] = log(((1 + total(%d))/termoccur(%d));\n",
 342                          i, rel->doc_frequency_vec[0],
 343                          rel->doc_frequency_vec[i]);
 344             wrbuf_printf(w, "%s: relevance += 100000 * tf[%d](%f) * "
 345                          "idf[%d](%f) (%d);\n",
 346                          e->display_str, i, termfreq, i, idfvec[i], add);
 347             relevance += add;
 348             e = e->next;
 349         }
 350         if (!rel->rank_cluster)
 351         {
 352             struct record *record;
 353             int cluster_size = 0;
 354
 355             for (record = rec->records; record; record = record->next)
 356                 cluster_size++;
 357
 358             wrbuf_printf(w, "score = relevance(%d)/cluster_size(%d);\n",
 359                          relevance, cluster_size);
 360             relevance /= cluster_size;
 361         }
 362         else
 363         {
 364             wrbuf_printf(w, "score = relevance(%d);\n", relevance);
 365         }
 366         rec->relevance_score = relevance;
 367     }
 368     reclist_leave(reclist);
 369     xfree(idfvec);
 370 }
 371
 372 /*
 373  * Local variables:
 374  * c-basic-offset: 4
 375  * c-file-style: "Stroustrup"
 376  * indent-tabs-mode: nil
 377  * End:
 378  * vim: shiftwidth=4 tabstop=8 expandtab
 379  */
 380