Tweak rank/relevance algorithm

[pazpar2-moved-to-github.git] / src / relevance.c

diff --git a/src/relevance.c b/src/relevance.c
index 7f1943a..b9fc0e1 100644 (file)
--- a/src/relevance.c
+++ b/src/relevance.c
@@ -34,6 +34,7 @@ struct relevance
     int vec_len;
     struct word_entry *entries;
     pp2_charset_token_t prt;
+    int rank_cluster;
     NMEM nmem;
 };
 
@@ -45,18 +46,23 @@ struct word_entry {
 };
 
 static int word_entry_match(struct word_entry *entries, const char *norm_str,
-                            const char *frank, int *local_mult)
+                            const char *rank, int *mult)
 {
     for (; entries; entries = entries->next)
     {
         if (!strcmp(norm_str, entries->norm_str))
         {
             const char *cp = 0;
-            if (frank && (cp = strchr(frank, ' ')))
+            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 - frank) == strlen(entries->ccl_field) &&
-                    memcmp(entries->ccl_field, frank, cp - frank) == 0)
-                    *local_mult = atoi(cp + 1);
+                if ((cp - rank) == strlen(entries->ccl_field) &&
+                    memcmp(entries->ccl_field, rank, cp - rank) == 0)
+                    *mult = atoi(cp + 1);
             }
             return entries->termno;
         }
@@ -65,8 +71,8 @@ static int word_entry_match(struct word_entry *entries, const char *norm_str,
 }
 
 void relevance_countwords(struct relevance *r, struct record_cluster *cluster,
-                          const char *words, int multiplier, const char *name,
-                          const char *frank)
+                          const char *words, const char *rank,
+                          const char *name)
 {
     int *mult = cluster->term_frequency_vec_tmp;
     const char *norm_str;
@@ -75,10 +81,11 @@ void relevance_countwords(struct relevance *r, struct record_cluster *cluster,
     for (i = 1; i < r->vec_len; i++)
         mult[i] = 0;
 
+    assert(rank);
     while ((norm_str = pp2_charset_token_next(r->prt)))
     {
-        int local_mult = multiplier;
-        int res = word_entry_match(r->entries, norm_str, frank, &local_mult);
+        int local_mult = 0;
+        int res = word_entry_match(r->entries, norm_str, rank, &local_mult);
         if (res)
         {
             assert(res < r->vec_len);
@@ -141,7 +148,8 @@ static void pull_terms(struct relevance *res, struct ccl_rpn_node *n)
 }
 
 struct relevance *relevance_create_ccl(pp2_charset_fact_t pft,
-                                       struct ccl_rpn_node *query)
+                                       struct ccl_rpn_node *query,
+                                       int rank_cluster)
 {
     NMEM nmem = nmem_create();
     struct relevance *res = nmem_malloc(nmem, sizeof(*res));
@@ -150,6 +158,7 @@ struct relevance *relevance_create_ccl(pp2_charset_fact_t pft,
     res->nmem = nmem;
     res->entries = 0;
     res->vec_len = 1;
+    res->rank_cluster = rank_cluster;
     res->prt = pp2_charset_token_create(pft, "relevance");
     
     pull_terms(res, query);
@@ -222,15 +231,9 @@ void relevance_prepare_read(struct relevance *rel, struct reclist *reclist)
             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
@@ -244,7 +247,17 @@ void relevance_prepare_read(struct relevance *rel, struct reclist *reclist)
         for (t = 1; t < rel->vec_len; t++)
         {
             float termfreq = (float) rec->term_frequency_vecf[t];
-            relevance += 100000 * (termfreq * idfvec[t] + 0.0000005);  
+            relevance += 100000 * termfreq * idfvec[t];
+        }
+        if (!rel->rank_cluster)
+        {
+            struct record *record;
+            int cluster_size = 0;
+
+            for (record = rec->records; record; record = record->next)
+                cluster_size++;
+            
+            relevance /= cluster_size;
         }
         rec->relevance_score = relevance;
     }