src/relevance.c

   1 /* This file is part of Pazpar2.
   2    Copyright (C) 2006-2012 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 struct relevance
  32 {
  33     int *doc_frequency_vec;
  34     int *term_frequency_vec_tmp;
  35     int *term_pos;
  36     int vec_len;
  37     struct word_entry *entries;
  38     pp2_charset_token_t prt;
  39     int rank_cluster;
  40     double follow_factor;
  41     double lead_decay;
  42     int length_divide;
  43     NMEM nmem;
  44 };
  45
  46 struct word_entry {
  47     const char *norm_str;
  48     const char *display_str;
  49     int termno;
  50     char *ccl_field;
  51     struct word_entry *next;
  52 };
  53
  54 static struct word_entry *word_entry_match(struct relevance *r,
  55                                            const char *norm_str,
  56                                            const char *rank, int *mult)
  57 {
  58     int i = 1;
  59     struct word_entry *entries = r->entries;
  60     for (; entries; entries = entries->next, i++)
  61     {
  62         if (*norm_str && !strcmp(norm_str, entries->norm_str))
  63         {
  64             const char *cp = 0;
  65             int no_read = 0;
  66             sscanf(rank, "%d%n", mult, &no_read);
  67             rank += no_read;
  68             while (*rank == ' ')
  69                 rank++;
  70             if (no_read > 0 && (cp = strchr(rank, ' ')))
  71             {
  72                 if ((cp - rank) == strlen(entries->ccl_field) &&
  73                     memcmp(entries->ccl_field, rank, cp - rank) == 0)
  74                     *mult = atoi(cp + 1);
  75             }
  76             return entries;
  77         }
  78     }
  79     return 0;
  80 }
  81
  82 void relevance_countwords(struct relevance *r, struct record_cluster *cluster,
  83                           const char *words, const char *rank,
  84                           const char *name)
  85 {
  86     int *mult = r->term_frequency_vec_tmp;
  87     const char *norm_str;
  88     int i, length = 0;
  89     double lead_decay = r->lead_decay;
  90     struct word_entry *e;
  91     WRBUF w = cluster->relevance_explain1;
  92
  93     pp2_charset_token_first(r->prt, words, 0);
  94     for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
  95     {
  96         mult[i] = 0;
  97         r->term_pos[i] = 0;
  98     }
  99
 100     assert(rank);
 101     while ((norm_str = pp2_charset_token_next(r->prt)))
 102     {
 103         int local_mult = 0;
 104         e = word_entry_match(r, norm_str, rank, &local_mult);
 105         if (e)
 106         {
 107             int res = e->termno;
 108             int j;
 109
 110             assert(res < r->vec_len);
 111             mult[res] += local_mult / (1 + log2(1 + lead_decay * length));
 112             wrbuf_printf(w, "%s: mult[%d] += local_mult(%d) / "
 113                          "(1+log2(1+lead_decay(%f) * length(%d)));\n",
 114                          e->display_str, res, local_mult, lead_decay, length);
 115             j = res - 1;
 116             if (j > 0 && r->term_pos[j])
 117             {
 118                 int d = length + 1 - r->term_pos[j];
 119                 mult[res] += mult[res] * r->follow_factor / (1 + log2(d));
 120                 wrbuf_printf(w, "%s: mult[%d] += mult[%d](%d) * follow(%f) / "
 121                              "(1+log2(d(%d));\n",
 122                              e->display_str, res, res, mult[res],
 123                              r->follow_factor, d);
 124             }
 125             for (j = 0; j < r->vec_len; j++)
 126                 r->term_pos[j] = j < res ? 0 : length + 1;
 127         }
 128         length++;
 129     }
 130
 131     for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
 132     {
 133         if (length == 0 || mult[i] == 0)
 134             continue;
 135         wrbuf_printf(w, "%s: field=%s vecf[%d] += mult[%d](%d)",
 136                      e->display_str, name, i, i, mult[i]);
 137         switch (r->length_divide)
 138         {
 139         case 0:
 140             wrbuf_printf(w, ";\n");
 141             cluster->term_frequency_vecf[i] += (double) mult[i];
 142             break;
 143         case 1:
 144             wrbuf_printf(w, " / log2(1+length(%d));\n", length);
 145             cluster->term_frequency_vecf[i] +=
 146                 (double) mult[i] / log2(1 + length);
 147             break;
 148         case 2:
 149             wrbuf_printf(w, " / length(%d);\n", length);
 150             cluster->term_frequency_vecf[i] += (double) mult[i] / length;
 151         }
 152         cluster->term_frequency_vec[i] += mult[i];
 153     }
 154
 155     cluster->term_frequency_vec[0] += length;
 156 }
 157
 158 static void pull_terms(struct relevance *res, struct ccl_rpn_node *n)
 159 {
 160     char **words;
 161     int numwords;
 162     char *ccl_field;
 163     int i;
 164
 165     switch (n->kind)
 166     {
 167     case CCL_RPN_AND:
 168     case CCL_RPN_OR:
 169     case CCL_RPN_NOT:
 170     case CCL_RPN_PROX:
 171         pull_terms(res, n->u.p[0]);
 172         pull_terms(res, n->u.p[1]);
 173         break;
 174     case CCL_RPN_TERM:
 175         nmem_strsplit(res->nmem, " ", n->u.t.term, &words, &numwords);
 176         for (i = 0; i < numwords; i++)
 177         {
 178             const char *norm_str;
 179
 180             ccl_field = nmem_strdup_null(res->nmem, n->u.t.qual);
 181
 182             pp2_charset_token_first(res->prt, words[i], 0);
 183             while ((norm_str = pp2_charset_token_next(res->prt)))
 184             {
 185                 struct word_entry **e = &res->entries;
 186                 while (*e)
 187                     e = &(*e)->next;
 188                 *e = nmem_malloc(res->nmem, sizeof(**e));
 189                 (*e)->norm_str = nmem_strdup(res->nmem, norm_str);
 190                 (*e)->ccl_field = ccl_field;
 191                 (*e)->termno = res->vec_len++;
 192                 (*e)->display_str = nmem_strdup(res->nmem, words[i]);
 193                 (*e)->next = 0;
 194             }
 195         }
 196         break;
 197     default:
 198         break;
 199     }
 200 }
 201
 202 struct relevance *relevance_create_ccl(pp2_charset_fact_t pft,
 203                                        struct ccl_rpn_node *query,
 204                                        int rank_cluster,
 205                                        double follow_factor, double lead_decay,
 206                                        int length_divide)
 207 {
 208     NMEM nmem = nmem_create();
 209     struct relevance *res = nmem_malloc(nmem, sizeof(*res));
 210     int i;
 211
 212     res->nmem = nmem;
 213     res->entries = 0;
 214     res->vec_len = 1;
 215     res->rank_cluster = rank_cluster;
 216     res->follow_factor = follow_factor;
 217     res->lead_decay = lead_decay;
 218     res->length_divide = length_divide;
 219     res->prt = pp2_charset_token_create(pft, "relevance");
 220
 221     pull_terms(res, query);
 222
 223     res->doc_frequency_vec = nmem_malloc(nmem, res->vec_len * sizeof(int));
 224     for (i = 0; i < res->vec_len; i++)
 225         res->doc_frequency_vec[i] = 0;
 226
 227     // worker array
 228     res->term_frequency_vec_tmp =
 229         nmem_malloc(res->nmem,
 230                     res->vec_len * sizeof(*res->term_frequency_vec_tmp));
 231
 232     res->term_pos =
 233         nmem_malloc(res->nmem, res->vec_len * sizeof(*res->term_pos));
 234
 235     return res;
 236 }
 237
 238 void relevance_destroy(struct relevance **rp)
 239 {
 240     if (*rp)
 241     {
 242         pp2_charset_token_destroy((*rp)->prt);
 243         nmem_destroy((*rp)->nmem);
 244         *rp = 0;
 245     }
 246 }
 247
 248 void relevance_newrec(struct relevance *r, struct record_cluster *rec)
 249 {
 250     if (!rec->term_frequency_vec)
 251     {
 252         int i;
 253
 254         // term frequency [1,..] . [0] is total length of all fields
 255         rec->term_frequency_vec =
 256             nmem_malloc(r->nmem,
 257                         r->vec_len * sizeof(*rec->term_frequency_vec));
 258         for (i = 0; i < r->vec_len; i++)
 259             rec->term_frequency_vec[i] = 0;
 260
 261         // term frequency divided by length of field [1,...]
 262         rec->term_frequency_vecf =
 263             nmem_malloc(r->nmem,
 264                         r->vec_len * sizeof(*rec->term_frequency_vecf));
 265         for (i = 0; i < r->vec_len; i++)
 266             rec->term_frequency_vecf[i] = 0.0;
 267     }
 268 }
 269
 270 void relevance_donerecord(struct relevance *r, struct record_cluster *cluster)
 271 {
 272     int i;
 273
 274     for (i = 1; i < r->vec_len; i++)
 275         if (cluster->term_frequency_vec[i] > 0)
 276             r->doc_frequency_vec[i]++;
 277
 278     r->doc_frequency_vec[0]++;
 279 }
 280
 281 // Prepare for a relevance-sorted read
 282 void relevance_prepare_read(struct relevance *rel, struct reclist *reclist)
 283 {
 284     int i;
 285     float *idfvec = xmalloc(rel->vec_len * sizeof(float));
 286
 287     reclist_enter(reclist);
 288     // Calculate document frequency vector for each term.
 289     for (i = 1; i < rel->vec_len; i++)
 290     {
 291         if (!rel->doc_frequency_vec[i])
 292             idfvec[i] = 0;
 293         else
 294         {
 295             /* add one to nominator idf(t,D) to ensure a value > 0 */
 296             idfvec[i] = log((float) (1 + rel->doc_frequency_vec[0]) /
 297                             rel->doc_frequency_vec[i]);
 298         }
 299     }
 300     // Calculate relevance for each document
 301     while (1)
 302     {
 303         int relevance = 0;
 304         WRBUF w;
 305         struct word_entry *e = rel->entries;
 306         struct record_cluster *rec = reclist_read_record(reclist);
 307         if (!rec)
 308             break;
 309         w = rec->relevance_explain2;
 310         wrbuf_rewind(w);
 311         for (i = 1; i < rel->vec_len; i++)
 312         {
 313             float termfreq = (float) rec->term_frequency_vecf[i];
 314             int add = 100000 * termfreq * idfvec[i];
 315
 316             wrbuf_printf(w, "idf[%d] = log(((1 + total(%d))/termoccur(%d));\n",
 317                          i, rel->doc_frequency_vec[0],
 318                          rel->doc_frequency_vec[i]);
 319             wrbuf_printf(w, "%s: relevance += 100000 * vecf[%d](%f) * "
 320                          "idf[%d](%f) (%d);\n",
 321                          e->display_str, i, termfreq, i, idfvec[i], add);
 322             relevance += add;
 323             e = e->next;
 324         }
 325         if (!rel->rank_cluster)
 326         {
 327             struct record *record;
 328             int cluster_size = 0;
 329
 330             for (record = rec->records; record; record = record->next)
 331                 cluster_size++;
 332
 333             wrbuf_printf(w, "score = relevance(%d)/cluster_size(%d);\n",
 334                          relevance, cluster_size);
 335             relevance /= cluster_size;
 336         }
 337         else
 338         {
 339             wrbuf_printf(w, "score = relevance(%d);\n", relevance);
 340         }
 341         rec->relevance_score = relevance;
 342     }
 343     reclist_leave(reclist);
 344     xfree(idfvec);
 345 }
 346
 347 /*
 348  * Local variables:
 349  * c-basic-offset: 4
 350  * c-file-style: "Stroustrup"
 351  * indent-tabs-mode: nil
 352  * End:
 353  * vim: shiftwidth=4 tabstop=8 expandtab
 354  */
 355