On the Structure and Organization of TREC data

Test Collections

x       query                   qrel    corpus
-----   ----------------------- ------  ------
task    routing         adhoc   adhoc   adhoc
-----   ----------------------- ------  ------
TREC1      1-50 (cd2)    51-100  51-100 cd12
TREC2     51-100(cd3)   101-150 101-150 cd12
TREC3    101-150(cd3)   151-200 151-200 cd12
TREC4                   201-250 201-250 cd23
TREC5                   251-300 251-300 cd24
TREC6                   301-350 301-350 cd45
TREC7                   351-400 351-400 cd45-cr
TREC8                   401-450 401-450 cd45-cr

Document Corpus

    wsj  WSJ  1987, 1988, 1989
    fr   FR   1989
    ap   AP   1989
    doe  DOE
    ziff ZF   1989, 1990
    wsj  WSJ  1990, 1991, 1992
    fr   FR   1988
    ap   AP   1988
    ziff ZF   1989, 1990
    sjm  SJM  1991
    ap   AP   1990
    pat  PT   1983-1991
    ziff ZF   1991, 1992
    ft   FT   1991-1994
    cr   CR   1993
    fr   FR   1994
    fbis FBIS 1996
    lat  LA   1989, 1990

Document Structure

These three seems to be around always. DOC, DOCNO, TEXT

A title shows up in many forms. TTL, TITLE, HEADLINE, H3, HT

Useful text blocks. SUMMARY

Some TEXT sections are strewn with funny comment tags and other tags too. 'within+' denotes such a TEXT section with one or more such tags within it.

TREC document structure table

            cd1     cd2     cd3      cd4       cd5
 ap         HEAD+   HEAD+   HEAD
 fr         within+ within+          within+
 wsj        HL      HL
 ziff       TITLE   TITLE   TITLE
                    SUMMARY SUMMARY
 patents                    TTL
 sjm                        LEADPARA
 cr                                  TTL
 ft                                  HEADLINE
 fbis                                         H3 (within+)
                                              HT (within+)
 la                                           HEADLINE

Query Structure

YEAR/TAG head num dom title desc smry narr con fac nat def  
1-100    x    x   x   x     x         x    x   x       x
101-150  x    x   x   x     x    x    x    x   x       x
151-200       x       x     x         x
201-250       x             x
251-300       x       x     x         x
301-350       x       x     x         x
351-400       x       x     x         x
401-450       x       x     x         x

Empty Documents

Depending on how you configure a search engine's parser (which tag contents to pick, etc.) documents may end up being empty, having no usable content. I usually make parsers as liberal as possible so that everything within a tag is consumed, but for contents and the mark-up tags themselves. The popular papers (and for that matter even recent ones) reporting experiments on TREC adhoc data don't mention details of the parser. TREC documents have a host of mark-up and little is known about why they were there, which search systems used them, or which TREC tasks were these annotations meant for.

So even after using the most liberal parser, there are some truly empty documents, and I have found 3 so far. Two of these fall prey to the tokenizer and stemmer;

File                DOCNO
cd1/doe/doe1_096    DOE1-96-1081
cd1/doe/doe2_013    DOE2-13-0573
cd1/doe/doe2_051    DOE2-51-1160


<DOCNO> DOE1-96-1081 </DOCNO>



<DOCNO> DOE2-13-0573 </DOCNO>


<DOCNO> DOE2-51-1160 </DOCNO>

Very Long Terms

Documents may have very long terms like this one from document LA072290-0141 in the CD5 LA Times sub-collection:


This is a name of a village in Wales; see the Wikipedia page about Llanfairpwllgwyngyll.

It is therefore recommended that you neither allocated just a small number of bytes for tokens or terms when building parsers, nor mistake such oddities as parser errors.