struct relevance
{
int *doc_frequency_vec;
+ int *term_frequency_vec_tmp;
int vec_len;
struct word_entry *entries;
pp2_charset_token_t prt;
+ int rank_cluster;
+ int follow_boost;
+ int lead_boost;
+ int length_divide;
NMEM nmem;
};
struct word_entry {
const char *norm_str;
+ const char *display_str;
int termno;
+ int follow_boost;
+ char *ccl_field;
struct word_entry *next;
};
-int word_entry_match(struct word_entry *entries, const char *norm_str)
+static int word_entry_match(struct relevance *r, const char *norm_str,
+ const char *rank, int *mult)
{
- for (; entries; entries = entries->next)
+ int i = 1;
+ struct word_entry *entries = r->entries;
+ for (; entries; entries = entries->next, i++)
{
- if (!strcmp(norm_str, entries->norm_str))
+ if (*norm_str && !strcmp(norm_str, entries->norm_str))
+ {
+ int extra = r->follow_boost;
+ struct word_entry *e_follow = entries;
+ const char *cp = 0;
+ int no_read = 0;
+ sscanf(rank, "%d%n", mult, &no_read);
+ rank += no_read;
+ while (*rank == ' ')
+ rank++;
+ if (no_read > 0 && (cp = strchr(rank, ' ')))
+ {
+ if ((cp - rank) == strlen(entries->ccl_field) &&
+ memcmp(entries->ccl_field, rank, cp - rank) == 0)
+ *mult = atoi(cp + 1);
+ }
+ (*mult) += entries->follow_boost;
+ while ((e_follow = e_follow->next) != 0 && extra > 0)
+ {
+ e_follow->follow_boost = extra--;
+ }
return entries->termno;
+ }
+ entries->follow_boost = 0;
}
return 0;
}
void relevance_countwords(struct relevance *r, struct record_cluster *cluster,
- const char *words, int multiplier, const char *name)
+ const char *words, const char *rank,
+ const char *name)
{
- int *mult = cluster->term_frequency_vec_tmp;
+ int *mult = r->term_frequency_vec_tmp;
const char *norm_str;
int i, length = 0;
+ int lead_mult = r->lead_boost;
+ struct word_entry *e;
+ WRBUF w = cluster->relevance_explain1;
pp2_charset_token_first(r->prt, words, 0);
- for (i = 1; i < r->vec_len; i++)
+ for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
+ {
mult[i] = 0;
+ e->follow_boost = 0;
+ }
+ assert(rank);
while ((norm_str = pp2_charset_token_next(r->prt)))
{
- int res = word_entry_match(r->entries, norm_str);
+ int local_mult = 0;
+ int res = word_entry_match(r, norm_str, rank, &local_mult);
if (res)
{
assert(res < r->vec_len);
- mult[res] += multiplier;
+ mult[res] += local_mult + lead_mult;
}
+ if (lead_mult > 0)
+ --lead_mult;
length++;
}
- for (i = 1; i < r->vec_len; i++)
+ for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
{
- if (length > 0) /* only add if non-empty */
+ if (length == 0 || mult[i] == 0)
+ continue;
+ wrbuf_printf(w, "%s: field=%s vecf[%d] += mult(%d)",
+ e->display_str, name, i, mult[i]);
+ switch (r->length_divide)
+ {
+ case 0:
+ wrbuf_printf(w, ";\n");
+ cluster->term_frequency_vecf[i] += (double) mult[i];
+ break;
+ case 1:
+ wrbuf_printf(w, " / log2(1+length(%d));\n", length);
+ cluster->term_frequency_vecf[i] +=
+ (double) mult[i] / log2(1 + length);
+ break;
+ case 2:
+ wrbuf_printf(w, " / length(%d);\n", length);
cluster->term_frequency_vecf[i] += (double) mult[i] / length;
+ }
cluster->term_frequency_vec[i] += mult[i];
}
{
char **words;
int numwords;
+ char *ccl_field;
int i;
switch (n->kind)
{
const char *norm_str;
+ ccl_field = nmem_strdup_null(res->nmem, n->u.t.qual);
+
pp2_charset_token_first(res->prt, words[i], 0);
while ((norm_str = pp2_charset_token_next(res->prt)))
{
e = &(*e)->next;
*e = nmem_malloc(res->nmem, sizeof(**e));
(*e)->norm_str = nmem_strdup(res->nmem, norm_str);
+ (*e)->ccl_field = ccl_field;
(*e)->termno = res->vec_len++;
+ (*e)->display_str = nmem_strdup(res->nmem, words[i]);
(*e)->next = 0;
}
}
}
struct relevance *relevance_create_ccl(pp2_charset_fact_t pft,
- NMEM nmem, struct ccl_rpn_node *query)
+ struct ccl_rpn_node *query,
+ int rank_cluster,
+ int follow_boost, int lead_boost,
+ int length_divide)
{
+ NMEM nmem = nmem_create();
struct relevance *res = nmem_malloc(nmem, sizeof(*res));
int i;
res->nmem = nmem;
res->entries = 0;
res->vec_len = 1;
+ res->rank_cluster = rank_cluster;
+ res->follow_boost = follow_boost;
+ res->lead_boost = lead_boost;
+ res->length_divide = length_divide;
res->prt = pp2_charset_token_create(pft, "relevance");
-
+
pull_terms(res, query);
res->doc_frequency_vec = nmem_malloc(nmem, res->vec_len * sizeof(int));
for (i = 0; i < res->vec_len; i++)
- res->doc_frequency_vec[i] = 0;
+ res->doc_frequency_vec[i] = 0;
+
+ // worker array
+ res->term_frequency_vec_tmp =
+ nmem_malloc(res->nmem,
+ res->vec_len * sizeof(*res->term_frequency_vec_tmp));
return res;
}
if (*rp)
{
pp2_charset_token_destroy((*rp)->prt);
+ nmem_destroy((*rp)->nmem);
*rp = 0;
}
}
r->vec_len * sizeof(*rec->term_frequency_vec));
for (i = 0; i < r->vec_len; i++)
rec->term_frequency_vec[i] = 0;
-
+
// term frequency divided by length of field [1,...]
rec->term_frequency_vecf =
nmem_malloc(r->nmem,
r->vec_len * sizeof(*rec->term_frequency_vecf));
for (i = 0; i < r->vec_len; i++)
rec->term_frequency_vecf[i] = 0.0;
-
- // for relevance_countwords (so we don't have to xmalloc/xfree)
- rec->term_frequency_vec_tmp =
- nmem_malloc(r->nmem,
- r->vec_len * sizeof(*rec->term_frequency_vec_tmp));
}
}
idfvec[i] = 0;
else
{
- // This conditional may be terribly wrong
- // It was there to address the situation where vec[0] == vec[i]
- // which leads to idfvec[i] == 0... not sure about this
- // Traditional TF-IDF may assume that a word that occurs in every
- // record is irrelevant, but this is actually something we will
- // see a lot
- if ((idfvec[i] = log((float) rel->doc_frequency_vec[0] /
- rel->doc_frequency_vec[i])) < 0.0000001)
- idfvec[i] = 1;
+ /* add one to nominator idf(t,D) to ensure a value > 0 */
+ idfvec[i] = log((float) (1 + rel->doc_frequency_vec[0]) /
+ rel->doc_frequency_vec[i]);
}
}
// Calculate relevance for each document
while (1)
{
- int t;
int relevance = 0;
+ WRBUF w;
+ struct word_entry *e = rel->entries;
struct record_cluster *rec = reclist_read_record(reclist);
if (!rec)
break;
- for (t = 1; t < rel->vec_len; t++)
+ w = rec->relevance_explain2;
+ wrbuf_rewind(w);
+ for (i = 1; i < rel->vec_len; i++)
+ {
+ float termfreq = (float) rec->term_frequency_vecf[i];
+ int add = 100000 * termfreq * idfvec[i];
+
+ wrbuf_printf(w, "idf[%d] = log(((1 + total(%d))/termoccur(%d));\n",
+ i, rel->doc_frequency_vec[0],
+ rel->doc_frequency_vec[i]);
+ wrbuf_printf(w, "%s: relevance += 100000 * vecf[%d](%f) * "
+ "idf[%d](%f) (%d);\n",
+ e->display_str, i, termfreq, i, idfvec[i], add);
+ relevance += add;
+ e = e->next;
+ }
+ if (!rel->rank_cluster)
+ {
+ struct record *record;
+ int cluster_size = 0;
+
+ for (record = rec->records; record; record = record->next)
+ cluster_size++;
+
+ wrbuf_printf(w, "score = relevance(%d)/cluster_size(%d);\n",
+ relevance, cluster_size);
+ relevance /= cluster_size;
+ }
+ else
{
- float termfreq = (float) rec->term_frequency_vecf[t];
- relevance += 100000 * (termfreq * idfvec[t] + 0.0000005);
+ wrbuf_printf(w, "score = relevance(%d);\n", relevance);
}
rec->relevance_score = relevance;
}