Greetings. We have the first homework prepared for the multiparty discourse
group.
The resources you will need are available on the web page:
http://www.cs.rochester.edu/research/trains/annotation
There is an annotation manual that contains our current proposal for an
annotation scheme (tentatively named DAMSL), based on our experience
working with the scheme sketched out at the Penn workshop.
We also have developed an annotation tool, DAT, for this scheme that should
greatly ease the annotation process, and which produces SGML format files
containing your annotations. You will need Perl 5.002 or higher and the
Perl/tk package. We have pointers to sites where you can download these
tools.
We have three short dialogs that we'd like everyone to annotate using the
tool. These are already segmented into utterance units. WE ASK THAT YOU
DON'T RESEGMENT THE DIALOGS THIS TIME, BUT DO THE BEST YOU CAN WITH HOW WE
SEGMENTED THEM. We plan a second round of homework before the workshop
where we will explore the segmentation issue.
1. The first is a short TRAINS dialog. Information on the task and
instructions given the participants are all available from the web page, as
well as the map that each participant had. This dialog has speech files
attached to it that can be played using DAT. We have found that you need to
listen to the speech in order to accurately annotate the TRAINS dialogs.
2. Next is a keyboard dialog from the Coconut corpus from
U. Pittsburgh. Information on the task and materials given to the
participants are available on the web page. There are no speech files for
this obviously.
3. Finally, we have a short dialog used in the homework assignments for the
Penn workshop. It is a discussion between a person and a travel agent
about airline fares. The speech files aren't available, but we found that
it wasn't difficult to annotate without them.
We are hoping to do a second round of annotation, so we ask each of you to
annotate the dialogs and return to us BEFORE CHRISTMAS.
* the three SGML files containing your annotation of each dialog
* notes on problems that you had with DAT, the annotation scheme, and/or
the manual.
We would then try to get the second round materials out to you in early
January to be completed by Jan 24. Its a very tight schedule, but would
greatly enhance the productivity of the workshop if we made it!
Note that DAMSL is NOT intended to provide each research project with all
the distinctions that they need to make. It is intended as a uniform
abstract layer that each group can then refine to capture distinctions that
they need for their research. For instance, we are currently defining a
refinement of the scheme that we will use for our analysis of the TRAINS
dialogs. But this will not become part of the DAMSL specification, rather
it will remain specific to the TRAINS project. But by following the DAMSL
standard, our annotations can be shared and reused by other groups.
So as long as DAMSL is compatible with the distinctions you'd like to make,
then we should be happy. Of course, if there are distinctions that everyone
would like to make, then we should consider adding them to the standard.
If possible, we would like to try a second round of experiments before the
workshop. So please try your best to get the annotations in by the 25th.
The actual annotation of the dialogs should not take long once you've
invested the time in studying the manual.
James Allen and Mark Core
{james,mcore}@cs.rochester.edu
Last change: 13 Dec 96 by mcoreClick here to send in comments or questions.