Following is a list of the reasons why single letters and numbers are extremely frequent in the index and examples of how WordDelimiterFilter breaks tokens into parts that get searched as phrase queries:

 1  parsing of apostrophes

1. can't  => "can" "t"
2. French:  "l'art" =>  "l" "art" [i]
3. 1900's => "1900"  "s"
4. O'Sullivan => "o" "sullivan"

2   Initials

1. J.P. Morgan => "j" "p" "morgan"

3   Abreviations

1. I.B.M => "i" "b" "m"

4   alphanumerics

1. 7^th  =>  "7" "th"
2. p2p => "p" "2" "p"
3. p1314 => "p" "1314"

5   punctuation in numbers

1. 1,200 => "1" "200"
2. 5.3    =>  "5" "3"
3. 7:30 => "7" "30"

6   hypens

1. 1950-56 => "1950" "56"
2. Two-Party => "Two" "Party"

Endnotes

[i] The are some specific issues with tokenizing French.  See:

• http://en.wikipedia.org/wiki/Elision_in_the_French_language
• http://mail-archives.apache.org/mod_mbox/lucene-java-user/200211.mbox/%3c5.1.0.14.0.20021121145130.00a97b70@mailbox.uottawa.ca%3e
• http://issues.apache.org/jira/browse/LUCENE-906
• http://blogs.msdn.com/correcteurorthographiqueoffice/archive/2005/12/07/500807.aspx
• http://en.wikipedia.org/wiki/Apostrophe_(mark)#Non-English_use