|
The Potsdam Dialogue Lab
We've recently begun to do experimental work on dialogue, mostly
within the DEAWU
project, but also in classes we teach. This page briefly introduces
our experiments, but mostly lists some tools (both third-party as well
as own-developments) and equipment we use, and resources we've built
(or are building).
Experiments
At a stretch, one might call the approach we're following in our
current series of experiments a Rutherfordian approach: we take a
dialogue situation and smash it against a wall, in the hope to learn
from the resulting smithereens what the parts are of that thing called
dialogue.
OK, that's not a great description, but we hope you get the
idea: We systematically disturb a conversational setting (for example
by introducing turn-taking restrictions or channel problems),
in order to learn from the strategies that the dialogue participants
use to deal with these disturbances. Watch this space for a list of
publications about the experiments.
Tools for experimental dialogue research
Recommended Third-Party Tools
Well, these tools hardly need our endorsement, but for completeness,
here's what we use to get from recording to analysis:
- To help with the transcription of recordings, and for acoustic
analysis, we
use Praat, the
premiere computational phonetics tool from the University of
Amsterdam. (Here's a
page with useful scripts to automatise some frequent tasks in Praat,
courtesy of Mietta Lennes of the University of Helsinki.)
-
For further annotations on the transcriptions (dialogue acts,
sentence types, co-reference, etc.), we use
MMAX, an
annotation editor built by EML research Heidelberg. We use it
because it allows for easy definition of annotation schemata,
flexible definition of visualisation styles, and is based
consistently on stand-off markup.
-
For the annotation of video material we use
Anvil (built
by M. Kipp from DFKI, Saarbrücken) which integrates quite nicely into
this toolchain, as it can import TextGrids from Praat.
Tools developed at Potsdam
The tools described above are great, but for some specialised
(support) tasks, we had to build some tools of our own, which we are
offering to the community here in the hope that they may be useful to
others as well.
-
ZeitWort is a tool for displaying all kinds of
interval-based textual information. This could be (transcriptions
or annotations of) utterances in a dialogue, which is what we use
it for, but it could also be events occuring during the execution
of a program. Whatever, as long as it has a start and end time,
we'll display it.
A very useful feature is the "fit to
page" mode which lets you view whole dialogues at a glance:
great to get a quick feel for the structure of a
dialogue. Zeitwort's display of interval-based data is fully
configurable; e.g., you can realise a display where turns are
shown as bars, with dialogue act annotations encoded by different
background colours.
(ZeitWort's main developer is Micha Brandt.)
This is an example where 5 kinds of information (speaker A
transcription and dialogue act annotation, same for speaker B,
plus move annotation) are layed out in one timeline.
This is an example of the overview mode, showing an excerpt of
a dialogue.
Zeitwort can be downloaded
from its
sourceforge page. If you try it out, please let us know how
you like it!
-
Less a proper tool, more a collection of
helpers: praat2mmax. Contains a) a script
that converts Praat TextGrids (e.g., transcriptions of dialogues) into
b) our default MMAX2-project style for dialogues,
which can form the basis for all kinds of dialogue annotations.
-
Also not a tool per se, and not even directly related to
experimental data collection, but I'll put it here anyways: The
Parrot Sketch, or: How to turn your computer into a parrot
using only freely available tools. Detailed instructions
that explain how to install tools like
Sphinx4
(automatic speech recognition),
freeTTS
(Text-to-Speech syntheziser)
and Dipper
(Java) (Dialogue Management framework), within
the OAA (Open Agent
Architecture), plus a few scripts that make this task easier.
We use this set-up as a basis for teaching dialogue management in
practical courses.
The instructions can be read
separately here, the full
package is available here.
Equipment in the dialogue lab
-
We've built a little hardware box that realises a full-duplex
communication channel (basically, it connects two headsets, and
also offers a couple of buttons and signal lights) that
is fully computer-controllable (using the LabJack U12 USB breakout
box). With this we can realise reaction-time experiments, things
like "push-to-talk" turn-taking policies, etc.
-
We also have available some digital video cameras for recording
participants in our experiments.
Resources built
As part of
the DEAWU project
we are running a series of experiments where the same communcation
situation is systematically distorted in a variety of ways. The
setting is described in detail in [tba]. We make the collected corpora
available here. A computer programme for supporting the setting will
be available here soon as well.
- The first kind of "distortion" was the introduction of a
"push-to-talk" turn-taking regime. Here we wanted to investigate the
effect of such restrictions on the structure of the resulting
dialogue. This work was mostly done by our student Tatjana Lucht, as
part of her diploma thesis.
[back]
[last changed - 04/08/2006]
|
|