For each line in the SEE results tables:

Document selector code (A-J)

    a letter code assigned to the person at NIST (or other
        evaluation: TREC or TDT) who chose the documents to be
        summarized

Model summarizer code (A-J)

    a letter code assigned to the person who created the model
    summary
    
Assessor code (A-J)

    a letter code assigned to the person who judged the peer
        summary

Peer summarizer code (baseline[1-5], manual[A-J], or system
submission[6-26])

        a code indicating the source of the peer summary, i.e., the
        one being judged.


    1-5 indicates a simple automatically algorithm was used. NIST
        created these and the definitions are in the guidelines

     A-J indicate a human created this summary - this way we can
        compare manual against manual to see an upper limit on    
        performance

     6-26 indicate that a participating system creating the summary
         As I indicated in the email announcing results, these codes
         are found in the table on the results page:

            <http://duc.nist.gov/duc2003/results.html>

         If you want lines for the summaries you submitted, just
         look at the table for the task you are interested in, and
         at all lines with a peer summarizer code equal to your code
         from the results page table.

Count of quality questions with non-0 answers

    Simply the number of the 12 quality questions with scores > 0

Mean of the quality question scores
    
    Average of the 12 scores each with value 0-3 

######### 12 peer quality questions ask for counts of ERRORS
Q1        (0 = 0, 1 = 1-5, 2 = 6-10, 3 = 11 or more)
Q2      (No questions asked on 10-word summaries)
Q3 ... Q12

         You can find these questions by following the link to the
         manual evaluation protocol. No quality questions were asked
     about the task 1 (very short summaries) since the format
         of those summaries was unrestricted and most of the questions
         need not have applied.

Fraction of unmarked peer units at least related to the model's subject

     We ask the assessor what fraction of the number of peer units
         which didn't overlap at all in meaning with any model unit was
         at least related to the subject of the model

Number of peer units

     Number of rough sentences in the peer

Number of marked peer units

         Number of peer units that the assessor felt expressed at least
         some of the meaning of the model

Number of unmarked peer units

         Number of peer units that the assessor felt did not express any
         of the meaning of the model

Number of model units

         The number of roughly elementary discourse units (e.g.,
         clauses etc) in the model

Mean coverage

         As indicated in the protocol, the assessor judges the coverage
         by the peer summary of each unit in the model. This is the mean
         of those coverage scores.
           
Median coverage

     median of the per-model-unit coverage scores

Sample std of coverage scores

     sample standard deviation of the per-model-unit coverage scores

Mean length-adjusted coverage
Median length-adjusted coverage
Sample std of adjusted coverage scores

        These are the analogous statistics but for a coverage score
        that emphasizes brevity as well as coverage. Details here:

        <http://duc.nist.gov/duc2002/covbrev.html>