Animating the Knowledge Radio

Most text analysis tools either pre-index texts or
work on smaller corpora in order to give users an
interactive environment where they can ask questions of
texts. High Performance Computing facilities provide
the opportunity to develop analytical tools that process
large amounts of data in real time and that lets us animate
analysis. This paper presents a hypothesis about
how animated analysis can take advantage of HPC facilities
to provide useful information to the user, especially
when the results of an analytical process are complex
visualizations. In this paper we will do three things:
• Discuss the outcomes of the April 2008 workshop
on the Digital Humanities and High Performance
Computing at McMaster University. In particular
we will draw attention to the opportunities for digital
humanists to develop a high resolution visualization
agenda.
• Demonstrate the Big See model for large-scale text
visualization and discuss animated visualizations.
• Describe a preliminary model for a streaming
Knowledge Radio.
1. Outcomes of the Digital Humanities and
HPC Workshop
In April 2008 SHARCNET, an HPC consortium in
Southern Ontario, hosted a workshop on bridging the
gap between digital humanities research and HPC facilities.
The workshop brought together digital humanities
researchers and HPC researchers from the SHARCNET
participating universities to discuss the challenges and
identify the opportunities for collaboration. The mission
of the workshop was to ask and begin answering the following
questions:
• What are the opportunities for the use of HPC facilities
for humanities research? • What examples are there of good practices and research
innovation?
• What are the barriers to humanists using HPC facilities
like SHARCNET and how can they be overcome?
• What concrete steps can SHARCNET (and by extension
other HPC facilities) take to reach out to
computing humanists and to then support their research?
In the spirit of open research the notes, action items, and
report at available online (See https://www.sharcnet.ca/
dh-hpc/index.php/Main_Page). Some of the key initiatives
that were recommended to SHARCNET were:
• Humanists need introductory documentation about
HPC and its uses in the humanities. John Bonnet,
working with Geoffrey Rockwell and Kyle Kuchmey
prepared an introduction to High Performance
Computing in the Arts and Humanities with some
examples. (See http://www.sharcnet.ca/Documents/
HHPC/hpcdh.html)
• For humanists to take advantage of HPC facilities
we need training opportunities, access to fellowships
that place us into HPC facilities, and opportunities
to prototype ideas with HPC folk (what we
call charettes).
• The Digital Humanities and other disciplines that
use HPC facilities should collaborate around information
visualization. The arts and humanities have
much to offer in the way of visual ideas and traditions
of interpreting the visual. Humanists likewise
can make good use of visualization facilities that are
developed by HPC facilities.
One of the key opportunities identified were in largescale
and/or high resolution visualization. How would
we represent evidence in the humanities if the size of
the display, the resolution of the display and the speed of
processing were not an issue? And that is the subject of
this paper - experiments in developing a model for largescale
representation.
2. The Big See and Animated Visualization
The Big See is an experiment in high-performance text
visualization. The project has developed a prototype
of how a text or corpus of texts could be represented if
processing and the resolution of the display were not an
issue. Most text visualizations, like word clouds and distribution
graphs, are designed for the personal computer
screen. In the Big See we anticipate wall displays with
3 dimensional display capabilities and the processing
to manipulate large amounts of data like all the content
words of a corpus in real time. The Big See proposes one
visual idea of what such a high performance visualization
would look like as it is generated and once it is manipulable.
To be clear, the current version of TBS does
not need to run on an HPC system, it can run on a PC,
but it models visual ideas for anticipated wall display
systems. The question we asked ourselves was:
How could we represent a text or corpus if we had high
resolution, wall-sized displays and processing was not an
issue?
The idea was to take a 2-dimensional visualization that
was based loosely on the successful Weighted Centroid
model that has been beautifully implemented by TextArc,
and to add a third dimension to it and real time
manipulation. The Big See in its default setting shows a
pipe of the 20 most frequently used words, each of which
is a line stretching the length of the pipe with markers
where that word occurs. The pipe of lines can be turned
in three dimensions so that it can be treated as a revolving
barrel of distribution graphs. In the center of the pipe
you have the text itself that recedes off into the distance
much like the opening text of the Star Wars movies.
Clicking on an instance of a word on the pipe advances
the text to the appropriate point. The Big See is currently
implemented as a PC application that has controls for the
various parameters, though the source was written to be
run on an HPC system.
Animated Visualization. One of the unanticipated outcomes
of the project was that we found the live generation
of the visualization compelling in and of itself. It
has the virtue that it makes the final visualization understandable
as you can see how, as the text is processed
(and marches off into the horizon), the high frequency
vocabulary changes. It also has the virtue that is animates
a computer reading of the text in a linear fashion
(starting with the first word and updating the visualization
word by word. Users can infer things about the vocabulary
of the text as it proceeds, though we have to be
careful not to confuse the computer reading (and animating)
a text with human reading, which is not necessarily
so linear. This leads us to hypothesize that the animation
of analytical processes can bear useful information for
users of text analysis tools if properly paced and if they
can represent the process. An animation can stand in for
a pragmatic demonstration of what the computer is doing
in its black box - “it reads in a word and adds it to
the line for that work moving the line towards the 12:00
noon spot on the centroid if the frequency surpasses that of another word.” Animations have, along with interactivity,
the prima facie capability to bring processes alive
giving users an intuitive understanding of what the final
visualization represents. Obviously, they also have the
ability to mislead the user, which suggests a fruitful avenue
for further research. What are the best practices in
information animation?
3. The Knowledge Radio
The Knowledge Radio is an extension of work done on
the Big See, but adds the following aspects:
• Instead of working with a large, static corpus, what
if we were to work with a large, open-ended or dynamic
corpus where new input would modify the visualization
as it was being processed (what Ben Fry
calls organic information visualization)?
• What might be the most useful types of information
to display to users for a dynamically analyzed diachronic
corpus?
An example use-case for the Knowlege Radio is a
blogger who wants to examine how a specific concept
evolves over time on the web. The blogger would provide
a search term or semantic field to a tool that would
begin querying a search engine by successive date ranges
and provide visual information on aggregate data as
it was crawling results. The blogger could fine-tune the
parameters of the visualization and scrub along a timeline
to replay certain moments, or to fast-forward to the
current point of analysis. The visualization might resemble
something like the “code_swarm” project, which
represents commit activity by individuals in several
open-source project (http://vis.cs.ucdavis.edu/ ~ogawa/
codeswarm/).
From a tool implementation perspective an interesting
challenge is to process a corpus as a stream rather than
as a static object, especially in ways that would permit
the user to playback segments, to compare segments, to
modify parameters for visualizing segments, and to summon
previously processed text for closer reading. This
requires a well planned model for maintaining and updating
aggregate data (as new text is processed) but also
for storing relevant data from previously processed text.
This portion of the presentation will focus on describing
the theoretical aspects and design principles of the
Knowlege Radio but will also briefly demonstrate the
state of the current Knowlege Radio prototype.
4. Links
Fry, Ben, “Organic Information Visualization” <benfry.
com/organic/>
Ogawa, Michael, “code_swarm” <vis.cs.ucdavis.
edu/~ogawa/codeswarm/>
Paley, Bradford, “TextArc” <www.textarc.org/>
Rockwell, Geoffrey et al. “The Big See” <tada.mcmaster.
ca/view/Main/BigSee>
SHARCNET: <www.sharcnet.ca/>