===============================================
	NXT SWITCHBOARD CORPUS PUBLIC RELEASE
			2008
===============================================


This is the public release of the NXT-format version of the Switchboard
Corpus. Before use, you MUST read 00LICENSE-SWBD.txt in this directory.

This file contains important information for users of the Switchboard
Corpus. It has some important notes concerning the annotations,
describing unusual features and approximated information.


------------------------------------------------------
	Getting started
------------------------------------------------------

If you have used NXT before, you can get stuck in with the metadata at
xml/swbd-metadata.xml.

You can find out about the corpus structure and how to use the query
language  by reading the website and README.SWBD-QUERIES.txt.

------------------------------------------------------
	Documentation
------------------------------------------------------

Most of the documentation for the release is at 
http://groups.inf.ed.ac.uk/switchboard/.

Also in this directory:
  README.SWBD-QUERIES.txt	Corpus structure and queries.
  README.EXERCISES.txt		Some rather old exercises to try on the data.
  README.TOOLS.txt		A guide to GUIS for the data.
  switchboard-guis.sh/.bat	Scripts to start the GUIs (read README.TOOLS.txt first).

------------------------------------------------------
	Syntactic/Phonetic trees
------------------------------------------------------

Phonetic annotations were constructed from the msstate
transcripts. They form a hierarchy phonwords--syllables--phonemes
where phonwords in the phonetic tree are parallel to terminals in 
the syntax tree. The correspondence is represented by pointers from
terminals to phonwords. This relationship can then be used in queries
to "jump" from one tree to the other. (See example queries.)

This mapping is necessary partly because of the differences in the
transcriptions (the phonetic tree comes from msstate and the syntactic
tree from Penn Treebank sources) and partly because the tokenisations
of words and of phonetic words/units diverge in the case of
contractions such as "don't".


------------------------------------------------------
	Approximated turn timings
------------------------------------------------------

Start and end times were calculated for the turns by looking at the
timings on the outermost of their underlying terminals. Where that
timing information is not available on the outer terminals, we
estimated it by reference to the closest available start and end times
of other turns by the same agent. In this case we also added attributes
for the "firstKnownStart" and "lastKnownEnd" times for the turn, which
represent the outermost timing information which is available from the
underlying terminals. Ther is also a new attribute "approx" in
every turn, which takes the values "true" or "false" according to
whether times were approximated.


------------------------------------------------------
	Disfluencies
------------------------------------------------------


The disfluency annotation present in the NXT format corpus is derived
from the Penn Treebank syntactic coding, not from the disfluencies
that form part of the dialogue act coding in its original format, and
were created using a pre-Treebank version of the corpus.
Parts of that disfluency coding, such as the non-sentence elements,
have no analogue in the NXT format at present.  We have not considered
what the relationship is between any Penn Treebank disfluencies (and
therefore the NXT format) and the ones from the dialogue act coding.
More information about the disfluency annotation that we didn't
include can be found in  "Dysfluency Annotation Stylebook for the
Switchboard Corpus", Marie Meteer, revised by Ann Taylor, 1995.
ftp://ftp.cis.upenn.edu/pub/treebank/swbd/doc/DFL-book.ps.

In the NXT format, a disfluency contains a reparandum and a repair,
and the terminals that make up these two elements are included
directly as nite:children. Often multiple repairs are made in a
disfluency, which results in a nested structure of disfluencies within
disfluencies. This nesting only takes the form of containment -
one disfluency will not partially overlap with another.