LAP, LICHEN, and DASS - Experiences combining data and tools

The Linguistic Atlas team at the University of
Georgia (LAP) and the LICHEN research team
at the University of Oulu have been investigating
the application of advances in information
engineering to humanities scholarship, in
particular methods for managing and mining
large-scale linguistic databases. Our aim has
been to bring together the linguistic and
technological expertise from Oulu and Georgia
in order to develop practical solutions for
common problems. In this paper we will show
the results of this cooperation--including the
Digital Archive of Southern Speech (DASS),
the pilot product for LAP-LICHEN released in
2009--and discuss our experiences and lessons
learned.
The LAP audio archive, amounting to 7000
hours of interviews, is an unparalleled resource
for study not only of the common language
of the US but for its culture more generally,
stories of daily life in America. Study of LAP
interviews so far has taken advantage only of
small bits of transcribed data extracted from
the full interview. The large, untranscribed bulk
of LAP interviews consists of the speakers’
accounts of their lives and their families, their
occupations and their diversions, their houses
and their land. Along with direct questioning
and conversational passages, in a quarter of the
thousands of audio files so far processed these
stories take the form of narratives of at least
one minute of continuous speech. To preserve
and to make this audio archive available puts
the people back into what has seemed to
some scholars to be a dry academic exercise
of collecting words. The LAP team has always
appreciated the personal, individual nature of
each interview, as well as the way that the
interviews can represent American culture; with
DASS, we can now share that appreciation
much more broadly with both the academic
community and with the public.
DASS is a collection of 64 interviews from
the Linguistic Atlas of the Gulf States (LAGS)
selected by LAGS Director Lee Pederson. Four
interviews come from each of the sixteen
regional LAGS sectors. Within each sector there
is one speaker from each Atlas Type: folk (largely
uneducated and insular), common (moderate
education and experience), cultivated (higher
education and/or participation in high culture).
One African American speaker was selected
from each sector, and folk, common, and
cultivated African American speakers are
distributed across the sectors. Speakers cover
a wide range of ages and social circumstances.
Over 400 hours of audio files are provided both
as large uncompressed .wav files (useful for
acoustic phonetic processing) and as thousands
of small .mp3 files for general listening.
Files are indexed according to subject matter
and speaker, according to a set list of 40
topics. Metadata and finding aids for particular
topics and kinds of speakers are provided,
including search tools and a GIS function.
Together the DASS data and the LICHEN tools
comprise about 200 GB of data, provided on
a portable USB drive. The interviews were
digitized and processed by the LAP team at
the University of Georgia with assistance from
a grant from the National Endowment for
the Humanities (PW-50007, “Digitization of
Atlas Audio Recordings”, with Opas-Hänninen
as partner). The first phase of the LICHEN
project was lead jointly by Opas-Hänninen and
Seppänen and funded by a grant from the
Emil Aaltonen Foundation (2006-2008, with
Kretzschmar as an international collaborator).
The University of Oulu and the University of
Georgia drew up legal agreements regarding
copyrights and the distribution of the software
with the data. The DASS/LICHEN package is
distributed by the LAP.

2
DASS is only the beginning, however. The
research team at the University of Oulu has
developed LICHEN as an electronic framework,
i.e. a type of toolbox, which handles multimodal
data. The toolbox has been developed using
two sets of data as testbeds, namely the
Oulu Archive of Minority Languages in the
North containing samples from the Kven,
Meänkieli, Veps and Karelian languages, as
well as DASS. Some of the data from minority
languages exists as video, which the toolbox
handles along with audio and text. We are
now working on a transcription tool, so that
audio and video materials can be provided
with textual representations aligned with the
sound and video. Finally, we are rebuilding
LICHEN as a Web-enabled framework, so that
users can access our language and cultural
materials remotely in line with the movement
for the creation of public corpora (see, e.g.,
Kretzschmar, Anderson, Beal, Corrigan, Opas-
Hänninen, and Plichta 2006; Kretzschmar,
Childs, and Dollinger 2009).
LAP-LICHEN cooperation began in 2004, when
Kretzschmar, Opas-Hänninen, Anderson, Beal,
and Corrigan met in Newcastle, to follow up
on conversations about public corpora begun
earlier. The group found that there were
common problems, standards, best practices
for the corpora managed by those attending,
and agreed to prepare a presentation at ICAME
in 2005 to highlight the possibility for shared
methods and joint actions (later published as the
2006 programmatic article). The LAP-LICHEN
collaboration bloomed as a result. Grants were
obtained for cooperation: the LICHEN model
was included in a large NEH proposal for
archival digital audio processing for LAP, and
conversely LAP was adopted as the large-
scale test for the enhancement of LICHEN
at Oulu. An operational version of LICHEN
that might be used for LAP was available and
demonstrated at DH2007 (Urbana). Further
development occurred in conjunction with
DH2008 (Oulu), leading to testing of LICHEN
on LAP materials in Georgia in Fall 2008.
As a result of these steps, the collaborators
rewrote the specifications for what the program
needed to do in February 2009, as substantial
bodies of archivally-processed LAP sound files
became available for LICHEN development and
testing. The collaborators decided that the key
requirements for the specification arose from
not only the characteristics of the data (e.g.
the structure consisting of interviews, reels and
clips with metadata and multimedia on each
level) and the desired uses of the framework
(e.g. search, browse, and view possible audio
selections on a map with GIS), but also from the
fact that both tools development and the final
stages of data preparation were taking place
simultaneously. The data had to be available as
flat files usable through any regular file browser,
and the sheer scale of the data ruled out the
possibility of creating duplicate files inside the
program structure as originally designed. The
software needed to make use of the existing files
and file structure, and so the task of combining
the data and the tools became an exercise in
conforming tools to data with as little effect
on the data itself as possible. To this end,
the collaborators developed two methods for
bringing in the data and its associated metadata:
1) a general-purpose parser that could traverse
file and folder structures and evaluate regular
expression patterns to parse metadata from the
file and folder names, and 2) a mechanism for
re-formatting standard spreadsheet documents
into XML documents for use by the tools. The
existence of the tools affected the preparation
of the data in two ways: 1) some incorrectly
named files and folders had to be renamed to
conform to the agreed format, and 2) all text
files had to be converted from the Microsoft
Word format into a plain text format for viewing
from within the developed tools. Both changes
were also beneficial to the data collection
itself, improving consistency within the data
and file support across different platforms. In
turn, the developed tools provide access into
the database through browsing of the data by
natural entities such as interviews, topics and
geographic location (as opposed to just files and
folders) and queries leveraging the full potential
of all the metadata fields.
The DASS product was launched in April 2009
at the SECOL conference (New Orleans), and
public distribution began in the summer of
2009 after resolution of the legal issues of
the collaboration with university authorities
at Georgia and Oulu. Even given the close
collaboration between developers, arriving at
legal language that suited the university
authorities proved to be quite difficult: the
Georgia lawyers thought they could be more
laissez-faire because the product was unlikely