/* This file is part of Pazpar2.
- Copyright (C) 2006-2011 Index Data
+ Copyright (C) 2006-2013 Index Data
Pazpar2 is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
#include "relevance.h"
#include "session.h"
+#include "client.h"
+#include "settings.h"
+
+#ifdef WIN32
+#define log2(x) (log(x)/log(2))
+#endif
struct relevance
{
int *doc_frequency_vec;
+ int *term_frequency_vec_tmp;
+ int *term_pos;
int vec_len;
struct word_entry *entries;
- pp2_relevance_token_t prt;
+ pp2_charset_token_t prt;
+ int rank_cluster;
+ double follow_factor;
+ double lead_decay;
+ int length_divide;
NMEM nmem;
+ struct norm_client *norm;
};
-
struct word_entry {
const char *norm_str;
+ const char *display_str;
int termno;
+ char *ccl_field;
struct word_entry *next;
};
-static void add_word_entry(NMEM nmem,
- struct word_entry **entries,
- const char *norm_str,
- int term_no)
+// Structure to keep data for norm_client scores from one client
+struct norm_client
{
- struct word_entry *ne = nmem_malloc(nmem, sizeof(*ne));
- ne->norm_str = nmem_strdup(nmem, norm_str);
- ne->termno = term_no;
-
- ne->next = *entries;
- *entries = ne;
+ int num; // number of the client
+ float max;
+ float min;
+ int count;
+ const char *native_score;
+ int scorefield;
+ float a,b; // Rn = a*R + b
+ struct client *client;
+ struct norm_client *next;
+ struct norm_record *records;
+};
+
+const int scorefield_none = -1; // Do not normalize anything, use tf/idf as is
+ // This is the old behavior, and the default
+const int scorefield_internal = -2; // use our tf/idf, but normalize it
+const int scorefield_position = -3; // fake a score based on the position
+
+// A structure for each (sub)record. There is one list for each client
+struct norm_record
+{
+ struct record *record;
+ float score;
+ struct record_cluster *clust;
+ struct norm_record *next;
+};
+
+// Find the norm_client entry for this client, or create one if not there
+struct norm_client *findnorm( struct relevance *rel, struct client* client)
+{
+ struct norm_client *n = rel->norm;
+ struct session_database *sdb;
+ while (n) {
+ if (n->client == client )
+ return n;
+ n = n->next;
+ }
+ n = nmem_malloc(rel->nmem, sizeof(struct norm_client) );
+ if ( rel->norm )
+ n->num = rel->norm->num +1;
+ else
+ n->num = 1;
+ n->count = 0;
+ n->max = 0.0;
+ n->min = 0.0;
+ n->client = client;
+ n->next = rel->norm;
+ rel->norm = n;
+ sdb = client_get_database(client);
+ n->native_score = session_setting_oneval(sdb, PZ_NATIVE_SCORE);
+ n->records = 0;
+ n->scorefield = scorefield_none;
+ yaz_log(YLOG_LOG,"Normalizing: Client %d uses '%s'", n->num, n->native_score );
+ if ( ! n->native_score || ! *n->native_score ) // not specified
+ n->scorefield = scorefield_none;
+ else if ( strcmp(n->native_score,"position") == 0 )
+ n->scorefield = scorefield_position;
+ else if ( strcmp(n->native_score,"internal") == 0 )
+ n->scorefield = scorefield_internal;
+ else
+ { // Get the field index for the score
+ struct session *se = client_get_session(client);
+ n->scorefield = conf_service_metadata_field_id(se->service, n->native_score);
+ }
+ yaz_log(YLOG_LOG,"Normalizing: Client %d uses '%s' = %d",
+ n->num, n->native_score, n->scorefield );
+ return n;
}
-int word_entry_match(struct word_entry *entries, const char *norm_str)
+// Add a record in the list for that client, for normalizing later
+static void setup_norm_record( struct relevance *rel, struct record_cluster *clust)
+{
+ struct record *record;
+ for (record = clust->records; record; record = record->next)
+ {
+ struct norm_client *norm = findnorm(rel, record->client);
+ struct norm_record *rp;
+ if ( norm->scorefield == scorefield_none)
+ break; // not interested in normalizing this client
+ rp = nmem_malloc(rel->nmem, sizeof(struct norm_record) );
+ norm->count ++;
+ rp->next = norm->records;
+ norm->records = rp;
+ rp->clust = clust;
+ rp->record = record;
+ if ( norm->scorefield == scorefield_position )
+ rp->score = 1.0 / record->position;
+ else if ( norm->scorefield == scorefield_internal )
+ rp->score = clust->relevance_score; // the tf/idf for the whole cluster
+ // TODO - Get them for each record, merge later!
+ else
+ {
+ struct record_metadata *md = record->metadata[norm->scorefield];
+ rp->score = md->data.fnumber;
+ assert(rp->score>0); // ###
+ }
+ yaz_log(YLOG_LOG,"Got score for %d/%d : %f ",
+ norm->num, record->position, rp->score );
+ if ( norm->count == 1 )
+ {
+ norm->max = rp->score;
+ norm->min = rp->score;
+ } else {
+ if ( rp->score > norm->max )
+ norm->max = rp->score;
+ if ( rp->score < norm->min && abs(rp->score) < 1e-6 )
+ norm->min = rp->score; // skip zeroes
+ }
+ }
+}
+
+// Calculate the squared sum of residuals, that is the difference from
+// normalized values to the target curve, which is 1/n
+static double squaresum( struct norm_record *rp, double a, double b)
{
- for (; entries; entries = entries->next)
+ double sum = 0.0;
+ for ( ; rp; rp = rp->next )
{
- if (!strcmp(norm_str, entries->norm_str))
- return entries->termno;
+ double target = 1.0 / rp->record->position;
+ double normscore = rp->score * a + b;
+ double diff = target - normscore;
+ sum += diff * diff;
+ }
+ return sum;
+}
+
+static void normalize_scores(struct relevance *rel)
+{
+ // For each client, normalize scores
+ struct norm_client *norm;
+ for ( norm = rel->norm; norm; norm = norm->next )
+ {
+ yaz_log(YLOG_LOG,"Normalizing client %d: scorefield=%d count=%d",
+ norm->num, norm->scorefield, norm->count);
+ norm->a = 1.0; // default normalizing factors, no change
+ norm->b = 0.0;
+ if ( norm->scorefield != scorefield_none &&
+ norm->scorefield != scorefield_position )
+ { // have something to normalize
+ double range = norm->max - norm->min;
+ int it = 0;
+ double a,b; // params to optimize
+ double as,bs; // step sizes
+ double chi;
+ // initial guesses for the parameters
+ if ( range < 1e-6 ) // practically zero
+ range = norm->max;
+ a = 1.0 / range;
+ b = abs(norm->min);
+ as = a / 3;
+ bs = b / 3;
+ chi = squaresum( norm->records, a,b);
+ while (it++ < 100) // safeguard against things not converging
+ {
+ // optimize a
+ double plus = squaresum(norm->records, a+as, b);
+ double minus= squaresum(norm->records, a-as, b);
+ if ( plus < chi && plus < minus )
+ {
+ a = a + as;
+ chi = plus;
+ }
+ else if ( minus < chi && minus < plus )
+ {
+ a = a - as;
+ chi = minus;
+ }
+ else
+ as = as / 2;
+ // optimize b
+ plus = squaresum(norm->records, a, b+bs);
+ minus= squaresum(norm->records, a, b-bs);
+ if ( plus < chi && plus < minus )
+ {
+ b = b + bs;
+ chi = plus;
+ }
+ else if ( minus < chi && minus < plus )
+ {
+ b = b - bs;
+ chi = minus;
+ }
+ else
+ bs = bs / 2;
+ yaz_log(YLOG_LOG,"Fitting it=%d: a=%f / %f b=%f / %f chi = %f",
+ it, a, as, b, bs, chi );
+ norm->a = a;
+ norm->b = b;
+ if ( abs(as) * 1000.0 < abs(a) &&
+ abs(bs) * 1000.0 < abs(b) )
+ break; // not changing much any more
+ }
+ }
+
+ if ( norm->scorefield != scorefield_none )
+ { // distribute the normalized scores to the records
+ struct norm_record *nr = norm->records;
+ for ( ; nr ; nr = nr->next ) {
+ double r = nr->score;
+ r = norm->a * r + norm -> b;
+ nr->clust->relevance_score = 10000 * r;
+ yaz_log(YLOG_LOG,"Normalized %f * %f + %f = %f",
+ nr->score, norm->a, norm->b, r );
+ // TODO - This keeps overwriting the cluster score in random order!
+ // Need to merge results better
+ }
+
+ }
+
+ } // client loop
+}
+
+
+static struct word_entry *word_entry_match(struct relevance *r,
+ const char *norm_str,
+ const char *rank, int *weight)
+{
+ int i = 1;
+ struct word_entry *entries = r->entries;
+ for (; entries; entries = entries->next, i++)
+ {
+ if (*norm_str && !strcmp(norm_str, entries->norm_str))
+ {
+ const char *cp = 0;
+ int no_read = 0;
+ sscanf(rank, "%d%n", weight, &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)
+ *weight = atoi(cp + 1);
+ }
+ return entries;
+ }
}
return 0;
}
-static struct word_entry *build_word_entries(pp2_relevance_token_t prt,
- NMEM nmem,
- const char **terms)
+int relevance_snippet(struct relevance *r,
+ const char *words, const char *name,
+ WRBUF w_snippet)
{
- int termno = 1; /* >0 signals THERE is an entry */
- struct word_entry *entries = 0;
- const char **p = terms;
+ int no = 0;
+ const char *norm_str;
+ int highlight = 0;
- for (; *p; p++)
+ pp2_charset_token_first(r->prt, words, 0);
+ while ((norm_str = pp2_charset_token_next(r->prt)))
{
- const char *norm_str;
+ size_t org_start, org_len;
+ struct word_entry *entries = r->entries;
+ int i;
- pp2_relevance_first(prt, *p, 0);
- while ((norm_str = pp2_relevance_token_next(prt)))
- add_word_entry(nmem, &entries, norm_str, termno);
- termno++;
+ pp2_get_org(r->prt, &org_start, &org_len);
+ for (; entries; entries = entries->next, i++)
+ {
+ if (*norm_str && !strcmp(norm_str, entries->norm_str))
+ break;
+ }
+ if (entries)
+ {
+ if (!highlight)
+ {
+ highlight = 1;
+ wrbuf_puts(w_snippet, "<match>");
+ no++;
+ }
+ }
+ else
+ {
+ if (highlight)
+ {
+ highlight = 0;
+ wrbuf_puts(w_snippet, "</match>");
+ }
+ }
+ wrbuf_xmlputs_n(w_snippet, words + org_start, org_len);
+ }
+ if (highlight)
+ wrbuf_puts(w_snippet, "</match>");
+ if (no)
+ {
+ yaz_log(YLOG_DEBUG, "SNIPPET match: %s", wrbuf_cstr(w_snippet));
}
- return entries;
+ return no;
}
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 *w = r->term_frequency_vec_tmp;
const char *norm_str;
int i, length = 0;
+ double lead_decay = r->lead_decay;
+ struct word_entry *e;
+ WRBUF wr = cluster->relevance_explain1;
+ int printed_about_field = 0;
- pp2_relevance_first(r->prt, words, 0);
- for (i = 1; i < r->vec_len; i++)
- mult[i] = 0;
+ pp2_charset_token_first(r->prt, words, 0);
+ for (e = r->entries, i = 1; i < r->vec_len; i++, e = e->next)
+ {
+ w[i] = 0;
+ r->term_pos[i] = 0;
+ }
- while ((norm_str = pp2_relevance_token_next(r->prt)))
+ assert(rank);
+ while ((norm_str = pp2_charset_token_next(r->prt)))
{
- int res = word_entry_match(r->entries, norm_str);
- if (res)
+ int local_weight = 0;
+ e = word_entry_match(r, norm_str, rank, &local_weight);
+ if (e)
{
+ int res = e->termno;
+ int j;
+
+ if (!printed_about_field)
+ {
+ printed_about_field = 1;
+ wrbuf_printf(wr, "field=%s content=", name);
+ if (strlen(words) > 50)
+ {
+ wrbuf_xmlputs_n(wr, words, 49);
+ wrbuf_puts(wr, " ...");
+ }
+ else
+ wrbuf_xmlputs(wr, words);
+ wrbuf_puts(wr, ";\n");
+ }
assert(res < r->vec_len);
- mult[res] += multiplier;
+ w[res] += local_weight / (1 + log2(1 + lead_decay * length));
+ wrbuf_printf(wr, "%s: w[%d] += w(%d) / "
+ "(1+log2(1+lead_decay(%f) * length(%d)));\n",
+ e->display_str, res, local_weight, lead_decay, length);
+ j = res - 1;
+ if (j > 0 && r->term_pos[j])
+ {
+ int d = length + 1 - r->term_pos[j];
+ wrbuf_printf(wr, "%s: w[%d] += w[%d](%d) * follow(%f) / "
+ "(1+log2(d(%d));\n",
+ e->display_str, res, res, w[res],
+ r->follow_factor, d);
+ w[res] += w[res] * r->follow_factor / (1 + log2(d));
+ }
+ for (j = 0; j < r->vec_len; j++)
+ r->term_pos[j] = j < res ? 0 : length + 1;
}
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 */
- cluster->term_frequency_vecf[i] += (double) mult[i] / length;
- cluster->term_frequency_vec[i] += mult[i];
+ if (length == 0 || w[i] == 0)
+ continue;
+ wrbuf_printf(wr, "%s: tf[%d] += w[%d](%d)", e->display_str, i, i, w[i]);
+ switch (r->length_divide)
+ {
+ case 0:
+ cluster->term_frequency_vecf[i] += (double) w[i];
+ break;
+ case 1:
+ wrbuf_printf(wr, " / log2(1+length(%d))", length);
+ cluster->term_frequency_vecf[i] +=
+ (double) w[i] / log2(1 + length);
+ break;
+ case 2:
+ wrbuf_printf(wr, " / length(%d)", length);
+ cluster->term_frequency_vecf[i] += (double) w[i] / length;
+ }
+ cluster->term_frequency_vec[i] += w[i];
+ wrbuf_printf(wr, " (%f);\n", cluster->term_frequency_vecf[i]);
}
cluster->term_frequency_vec[0] += length;
}
-struct relevance *relevance_create(pp2_charset_t pct,
- NMEM nmem, const char **terms)
+static void pull_terms(struct relevance *res, struct ccl_rpn_node *n)
{
- struct relevance *res = nmem_malloc(nmem, sizeof(struct relevance));
- const char **p;
+ char **words;
+ int numwords;
+ char *ccl_field;
int i;
- for (p = terms, i = 0; *p; p++, i++)
- ;
- res->vec_len = ++i;
- res->doc_frequency_vec = nmem_malloc(nmem, res->vec_len * sizeof(int));
- memset(res->doc_frequency_vec, 0, res->vec_len * sizeof(int));
+ switch (n->kind)
+ {
+ case CCL_RPN_AND:
+ case CCL_RPN_OR:
+ case CCL_RPN_NOT:
+ case CCL_RPN_PROX:
+ pull_terms(res, n->u.p[0]);
+ pull_terms(res, n->u.p[1]);
+ break;
+ case CCL_RPN_TERM:
+ nmem_strsplit(res->nmem, " ", n->u.t.term, &words, &numwords);
+ for (i = 0; i < numwords; i++)
+ {
+ 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)))
+ {
+ struct word_entry **e = &res->entries;
+ while (*e)
+ 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;
+ }
+ }
+ break;
+ default:
+ break;
+ }
+}
+void relevance_clear(struct relevance *r)
+{
+ if (r)
+ {
+ int i;
+ for (i = 0; i < r->vec_len; i++)
+ r->doc_frequency_vec[i] = 0;
+ }
+}
+
+struct relevance *relevance_create_ccl(pp2_charset_fact_t pft,
+ struct ccl_rpn_node *query,
+ int rank_cluster,
+ double follow_factor, double lead_decay,
+ int length_divide)
+{
+ NMEM nmem = nmem_create();
+ struct relevance *res = nmem_malloc(nmem, sizeof(*res));
+
res->nmem = nmem;
- res->prt = pp2_relevance_tokenize(pct);
- res->entries = build_word_entries(res->prt, nmem, terms);
+ res->entries = 0;
+ res->vec_len = 1;
+ res->rank_cluster = rank_cluster;
+ res->follow_factor = follow_factor;
+ res->lead_decay = lead_decay;
+ res->length_divide = length_divide;
+ res->norm = 0;
+ res->prt = pp2_charset_token_create(pft, "relevance");
+
+ pull_terms(res, query);
+
+ res->doc_frequency_vec = nmem_malloc(nmem, res->vec_len * sizeof(int));
+
+ // worker array
+ res->term_frequency_vec_tmp =
+ nmem_malloc(res->nmem,
+ res->vec_len * sizeof(*res->term_frequency_vec_tmp));
+
+ res->term_pos =
+ nmem_malloc(res->nmem, res->vec_len * sizeof(*res->term_pos));
+
+ relevance_clear(res);
return res;
}
{
if (*rp)
{
- pp2_relevance_token_destroy((*rp)->prt);
+ pp2_charset_token_destroy((*rp)->prt);
+ nmem_destroy((*rp)->nmem);
*rp = 0;
}
}
-void relevance_newrec(struct relevance *r, struct record_cluster *rec)
+void relevance_mergerec(struct relevance *r, struct record_cluster *dst,
+ const struct record_cluster *src)
{
- if (!rec->term_frequency_vec)
- {
- int i;
+ int i;
- // term frequency [1,..] . [0] is total length of all fields
- rec->term_frequency_vec =
- nmem_malloc(r->nmem,
- 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));
- }
+ for (i = 0; i < r->vec_len; i++)
+ dst->term_frequency_vec[i] += src->term_frequency_vec[i];
+
+ for (i = 0; i < r->vec_len; i++)
+ dst->term_frequency_vecf[i] += src->term_frequency_vecf[i];
}
+void relevance_newrec(struct relevance *r, struct record_cluster *rec)
+{
+ int i;
+
+ // term frequency [1,..] . [0] is total length of all fields
+ rec->term_frequency_vec =
+ nmem_malloc(r->nmem,
+ 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;
+}
void relevance_donerecord(struct relevance *r, struct record_cluster *cluster)
{
r->doc_frequency_vec[0]++;
}
+
+
// Prepare for a relevance-sorted read
void relevance_prepare_read(struct relevance *rel, struct reclist *reclist)
{
float *idfvec = xmalloc(rel->vec_len * sizeof(float));
reclist_enter(reclist);
+
// Calculate document frequency vector for each term.
for (i = 1; i < rel->vec_len; i++)
{
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);
+ wrbuf_puts(w, "relevance = 0;\n");
+ for (i = 1; i < rel->vec_len; i++)
{
- float termfreq;
-#if 1
- termfreq = (float) rec->term_frequency_vecf[t];
-#else
- if (rec->term_frequency_vec[0])
- {
- termfreq = (float)
- rec->term_frequency_vec[t] / rec->term_frequency_vec[0] ;
- }
- else
- termfreq = 0.0;
-#endif
- relevance += 100000 * (termfreq * idfvec[t] + 0.0000005);
+ 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 * tf[%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
+ {
+ wrbuf_printf(w, "score = relevance(%d);\n", relevance);
}
rec->relevance_score = relevance;
- }
+
+ // Build the normalizing structures
+ // List of (sub)records for each target
+ setup_norm_record( rel, rec );
+
+ // TODO - Loop again, merge individual record scores into clusters
+ // Can I reset the reclist, or can I leave and enter without race conditions?
+
+ } // cluster loop
+
+ normalize_scores(rel);
+
reclist_leave(reclist);
xfree(idfvec);
+
}
/*