/* Copyright (c) 2008 - Chris Buckley. Permission is granted for use and modification of this file for research, non-commercial purposes. */ #include "common.h" #include "sysfunc.h" #include "trec_eval.h" #include "trec_format.h" #include /* Read all relevance preference information from text_prefs_file. Preferences of user(s) for docs for a given qid is determined from text_prefs_file, which consists of text tuples of the form qid ujg ujsubg docno rel_level giving TREC document numbers (docno, a string) and their relevance level (rel_level,a non-negative float) to query qid (a string) for a user judgment sub-group (ujsubg, a string) within a user judgment group (ujg, a string). Fields are separated by whitespace, string fields can contain no whitespace. File may contain no NULL characters. Preferences are indicated indirectly by comparing rel_level of different docnos within the same user judgment sub group(JSG). A judgment sub group establishes preferences between all docnos with non-tied rel_levels within the group. Except possibly for 0.0, the actual values of rel_level are ignored by default; they only serve to establish a ranking within the JSG. If a user only expresses a preference between two docs, then that user JSG will have 2 lines in text_prefs_file: qid1 ujg1 sub1 docno1 3.0 qid1 ujg1 sub1 docno2 2.0 If a user completely ranks some small number N (5-10) of docs, then N lines are used. For example: qid1 ujg1 sub1 docno1 3.0 qid1 ujg1 sub1 docno2 2.0 qid1 ujg1 sub1 docno3 0.0 qid1 ujg1 sub1 docno4 6.0 qid1 ujg1 sub1 docno5 0.0 qid1 ujg1 sub1 docno6 2.0 establishes a total of 13 preferences (5 with docno4 preferred, 4 with docno1 preferred, 2 each with docno2 and docno6 preferred). If a given user has multiple preferences that aren't complete, the preferences are expressed in multiple JSGs within a single JG. For example: qid1 ujg1 sub1 docno1 3.0 qid1 ujg1 sub1 docno2 2.0 qid1 ujg1 sub1 docno3 1.0 qid1 ujg1 sub2 docno1 2.0 qid1 ujg1 sub2 docno2 1.0 qid1 ujg1 sub2 docno4 3.0 expressses 5 preferences (1>2, 1>3, 2 > 3, 4>1, 4>2). Note the duplicate 1 > 2 is not counted as a separate preference. A conventional pairwise preference file with no transistivity could be converted into this form, with two entries per JSG, the preferred doc with a rel of 2.0 and the non-preferred doc with a rel of 1.0. Multiple users are indicated by different JGs. For example: qid1 ujg1 sub1 docno1 3.0 qid1 ujg1 sub1 docno2 2.0 qid1 ujg2 sub1 docno1 0.0 qid1 ujg2 sub1 docno3 6.0 qid1 ujg2 sub1 docno4 2.0 qid1 ujg2 sub2 docno1 0.0 qid1 ujg2 sub2 docno2 8.0 expressses 5 preferences (1>2, 3>1, 4>1, 3>4, 2>1). A Judgment Group (JG) conceptually represents preferences for a single information need of a user at a single time. Within a single JG, it is an error if there are inconsistencies (doc A > doc B in one JSG, but B > A or B == A in another). The different JSGs within a JG are just a mechanism tha allows expressing partial ordering within a JG. Within a single JG, preferences are transistive: qid1 ujg1 sub1 docno1 3.0 qid1 ujg1 sub1 docno2 2.0 qid1 ujg1 sub1 docno3 1.0 qid1 ujg1 sub2 docno2 5.0 qid1 ujg1 sub2 docno4 4.0 expresses 5 preferences (1>2, 1>3, 2>3, 2>4, 1>4). There is no preference expressed between 3 and 4. Different JGs may contain contradictory preferences, as in an earlier example. These disagreements are realistic and desirable: users (or even the same user at different times) often do not agree with each other's preferences. Individual preference evaluation measures will handle these contradictions (or confirmations) in different ways. A rel_level of 0.0 by convention means that doc is non-relevant to the topic (in that user's opinion). it is an inconsistency (and an error) if a doc is assigned a rel_level of 0.0 in one JSG, but a different rel_level value in another JSG of the same JG. Some preference evaluation measures may handle 0.0 differently. Thus when converting a preference file in some other format into text_prefs format, do not assign a rel_level of 0.0 to a docno unless it is known that docno was considered nonrelevant. Handling of rel_level 0.0 separately addresses the general problem that the number of nonrelevant docs judged for a topic can be critical to fair evaluation - adding a couple of hundred preferences involving nonrelevant docs (out of the possibly millions or billions in a collection) can both change the importance of the topic when averaging and even change whether system A scores better than system B on a topic (even given identical retrieval on the added nonrel docs). How to handle this correctly for preference evaluation will be an important future research problem. */ static int parse_prefs_line (char **start_ptr, char **qid_ptr, char **jg_ptr, char **jsg_ptr, char **docno_ptr, char **rel_ptr); static int comp_lines_qid_docno (); /* static pools of memory, allocated here and never changed. Declared static so one day I can write a cleanup procedure to free them */ static char *trec_prefs_buf = NULL; static TEXT_PREFS_INFO *text_info_pool = NULL; static TEXT_PREFS *text_prefs_pool = NULL; static REL_INFO *rel_info_pool = NULL; /* Temp structure for values in input line */ typedef struct { char *qid; char *jg; char *jsg; char *docno; char *rel; } LINES; int te_get_prefs (EPI *epi, char *text_prefs_file, ALL_REL_INFO *all_rel_info) { int fd; int size = 0; char *ptr; char *current_qid; long i; LINES *lines; LINES *line_ptr; long num_lines; long num_qid; /* current pointers into static pools above */ REL_INFO *rel_info_ptr; TEXT_PREFS_INFO *text_info_ptr; TEXT_PREFS *text_prefs_ptr; /* Read entire file into memory */ if (-1 == (fd = open (text_prefs_file, 0)) || 0 >= (size = lseek (fd, 0L, 2)) || NULL == (trec_prefs_buf = malloc ((unsigned) size+2)) || -1 == lseek (fd, 0L, 0) || size != read (fd, trec_prefs_buf, size) || -1 == close (fd)) { fprintf (stderr, "trec_eval.get_prefs: Cannot read prefs file '%s'\n", text_prefs_file); return (UNDEF); } /* Append ending newline if not present, Append NULL terminator */ if (trec_prefs_buf[size-1] != '\n') { trec_prefs_buf[size] = '\n'; size++; } trec_prefs_buf[size] = '\0'; /* Count number of lines in file */ num_lines = 0; for (ptr = trec_prefs_buf; *ptr; ptr = index(ptr,'\n')+1) num_lines++; /* Get all lines */ if (NULL == (lines = Malloc (num_lines, LINES))) return (UNDEF); line_ptr = lines; ptr = trec_prefs_buf; while (*ptr) { if (UNDEF == parse_prefs_line (&ptr, &line_ptr->qid, &line_ptr->jg, &line_ptr->jsg, &line_ptr->docno, &line_ptr->rel)) { fprintf (stderr, "trec_eval.get_prefs: Malformed line %ld\n", (long) (line_ptr - lines + 1)); return (UNDEF); } line_ptr++; } num_lines = line_ptr-lines; /* Sort all lines by qid, then docno */ qsort ((char *) lines, (int) num_lines, sizeof (LINES), comp_lines_qid_docno); /* Go through lines and count number of qid */ num_qid = 1; for (i = 1; i < num_lines; i++) { if (strcmp (lines[i-1].qid, lines[i].qid)) /* New query */ num_qid++; } /* Allocate space for queries */ if (NULL == (rel_info_pool = Malloc (num_qid, REL_INFO)) || NULL == (text_info_pool = Malloc (num_qid, TEXT_PREFS_INFO)) || NULL == (text_prefs_pool = Malloc (num_lines, TEXT_PREFS))) return (UNDEF); rel_info_ptr = rel_info_pool; text_info_ptr = text_info_pool; text_prefs_ptr = text_prefs_pool; /* Go through lines and store all info */ current_qid = ""; for (i = 0; i < num_lines; i++) { if (strcmp (current_qid, lines[i].qid)) { /* New query. End old query and start new one */ if (i != 0) { text_info_ptr->num_text_prefs = text_prefs_ptr - text_info_ptr->text_prefs; text_info_ptr++; rel_info_ptr++; } current_qid = lines[i].qid; text_info_ptr->text_prefs = text_prefs_ptr; *rel_info_ptr = (REL_INFO) {current_qid, "prefs", text_info_ptr}; } text_prefs_ptr->jg = lines[i].jg; text_prefs_ptr->jsg = lines[i].jsg; text_prefs_ptr->rel_level = atof (lines[i].rel); text_prefs_ptr->docno = lines[i].docno; text_prefs_ptr++; } /* End last qid */ text_info_ptr->num_text_prefs = text_prefs_ptr - text_info_ptr->text_prefs; all_rel_info->num_q_rels = num_qid; all_rel_info->rel_info = rel_info_pool; Free (lines); return (1); } static int comp_lines_qid_docno (LINES *ptr1, LINES *ptr2) { int cmp = strcmp (ptr1->qid, ptr2->qid); if (cmp) return (cmp); return (strcmp (ptr1->docno, ptr2->docno)); } static int parse_prefs_line (char **start_ptr, char **qid_ptr, char**jg_ptr, char **jsg_ptr, char **docno_ptr, char **rel_ptr) { char *ptr = *start_ptr; /* Get qid */ while (*ptr != '\n' && isspace (*ptr)) ptr++; *qid_ptr = ptr; while (! isspace (*ptr)) ptr++; if (*ptr == '\n') return (UNDEF); *ptr++ = '\0'; /* Get Judgment Group */ while (*ptr != '\n' && isspace (*ptr)) ptr++; *jg_ptr = ptr; while (! isspace (*ptr)) ptr++; if (*ptr == '\n') return (UNDEF); *ptr++ = '\0'; /* Get Judgment Sub Group */ while (*ptr != '\n' && isspace (*ptr)) ptr++; *jsg_ptr = ptr; while (! isspace (*ptr)) ptr++; if (*ptr == '\n') return (UNDEF); *ptr++ = '\0'; /* Get docno */ while (*ptr != '\n' && isspace (*ptr)) ptr++; *docno_ptr = ptr; while (! isspace (*ptr)) ptr++; if (*ptr == '\n') return (UNDEF); *ptr++ = '\0'; /* Get relevance */ while (*ptr != '\n' && isspace (*ptr)) ptr++; if (*ptr == '\n') return (UNDEF); *rel_ptr = ptr; while (! isspace (*ptr)) ptr++; if (*ptr != '\n') { *ptr++ = '\0'; while (*ptr != '\n' && isspace (*ptr)) ptr++; if (*ptr != '\n') return (UNDEF); } *ptr++ = '\0'; *start_ptr = ptr; return (0); } int te_get_prefs_cleanup () { if (trec_prefs_buf != NULL) { Free (trec_prefs_buf); trec_prefs_buf = NULL; } if (text_info_pool != NULL) { Free (text_info_pool); text_info_pool = NULL; } if (text_prefs_pool != NULL) { Free (text_prefs_pool); text_prefs_pool = NULL; } if (rel_info_pool != NULL) { Free (rel_info_pool); rel_info_pool = NULL; } return (1); }