cluster_summary.txt

See .doc for abstract version.

14:45 - 16:15
11:00 - 12:00 [a moins]
arg. Some time in the afternoon or so
11:15 - 12:30
19:30 - 20:10
20:30 - 21:10
Clustering of cochlear implant users using Levenshtein distance
In ascending order of detail, written in Word (somehow)

Clustering of cochlear implant users with Levenshtein distance (SB
Chin, NC Sanders, in progress):
-- Why --
Building on previous work (Sanders & Chin, submitted) that found
Levenshtein distance to be comparable to human judgments of
intelligibility, clustering of cochlear implant users uses additional
information obtainable from this measure. Intelligibility judgments
are typically collected from adult speakers of English with no hearing
problems. Levenshtein distance correlates well with this measure, but
because of its computational nature can also calculate intelligibility
between any pair of cochlear implant users. This distance is new
information, because actual intelligibility judgments between two
implants users have not been obtained.

These clusters of cochlear implant users are the first step toward
automatic grouping of implant users by their phonological
characteristics. This could eventually allow computers to
automatically group a new implant user with an existing cluster and
thereby determine a system of phonological training.  If there are
differing effects of deafness on cochlear implant that result in
phonological differences, then clustering should be able to sort them
into groups [given enough distance information]. These groups can then
be analysed separately to determine the specific errors they make in
speech and how they should be corrected.

ASIDE: An important point
is that we are not looking at a priori factors for clustering
(most obvious example is total communication/oral
communication).

-- What --
The variant of Levenshtein distance that has been shown to correlate
best with human intelligibility judgments simply counts the number of
features that differ between the baseline speaker and a cochlear
implant user. Although intelligibility judgments can be gathered
easily from humans, with no transcription step required, Levenshtein
distance can also count the number of features that differ between two
cochlear implant users. This means that Levenshtein distance can be
measured between two cochlear implant users as easily as it can
from the baseline.

ASIDE: Because of the high correlation with human judgments from the
baseline, it seems reasonable to assume that Levenshtein distance
between two cochlear implant users correlates well with the
intelligibility judgments between the two, although these have not
been collected for comparison.

Thus, the clustering technique uses information not available to
previous percent-correct distances: comparisons between all implant
users. With just baseline scores, all intelligibility scores of the
same value cannot be differentiated. For example, three
intelligibility scores of 78% cannot be distinguished as different,
even if the actual phonologies of the speakers are quite
different. However, if you know that the intelligibility between
speaker 1 and speaker 2 is 80%, while the intelligibility between both
and speaker 3 is only 30%, then you have a better basis for grouping
speakers 1 and 2 together and putting speaker 3 into a separate category.

-Graph here so nobody has to read the gory details-

-- How --
The clustering technique used here is hierarchical binary
clustering. With this technique, every speaker is initially put alone
into a group. Then the two closest groups are merged. When there is
only one group left, the process is complete. This merging process
creates a tree that contains all speakers; clusters emerge from the
tree by how well the nested groups cohere. Currently, two speakers of
a set of eight have been found to cohere into a group outlying from
the rest, which cluster more closely around the baseline
speaker. Research is ongoing to determine what makes these two
speakers differ from the others.

TECHNICAL NOTE: The two closest groups are defined by using the
greatest distance between any member in the groups. This is called
`complete link' set distance.

ASIDE: I never figured out how to work in the term 'high-dimensional
triangulation' in a believeable way. This is too bad, because it
sounds really cool.