One
basic strategy is to seek an M is shared by all (or most) tokens which include S. This
is the first basic strategy employed in this paper. This can be done using direct evidence for
the presence of meaning M in each case, or indirect evidence, such as the prevalence of S in
tokens serving functional or positional roles which correlate with M.
However this is not easy, because every utterance means many things at many levels
(Schiffrin 1987; Traum 2000; Louwerse & Mitchell 2003). For example, the umm in Example
1, which was presented as indicating that the speaker was thinking, might also be interpreted
as meaning that he was withdrawing, or becoming serious, or wanting to slow the pace of the
interaction, or foreshadowing the imminent discussion of something significant, or showing
a polite reluctance to dominate the conversation, or cuing the other to listen closely, or
holding the floor, or hiding something, and so on. In past, sophisticated studies of some such
functions at various levels have been done, and there are a number of useful frameworks for
analysis. These, however, are mostly limited in that they focus on one level or one type of
function. There are, for example, studies which consider some non-lexical items as discourse
particles, connectors, acknowledgements, continuers, assessments, turn-taking cues, and so
on. However, as Fischer (2000) notes, these items ‘actually form a more homogeneous group
than suggested by the number of different descriptive labels’. For this reason the analysis
here was not done within any specific theory or framework; rather the shared meanings were
sought bottom-up, by observing similarities across the corpus.
The task of the analyst is to examine the entire set of tokens containing S, and pick out
the ‘best’ meaning, that is, the meaning component M which is (most) common across the
set. While examining the data various possible Ms were kept constantly in mind, namely
those identified as important in previous studies of conversation, non-verbal communication,
and inter-personal interaction. These include various functions involving discourse structure
marking, signaling of turn-taking intentions, negotiating agreement, signaling recognition and
comprehension, managing interpersonal relations such as control and affiliation, and express-
ing emotion, attitude, and affect.
For lack of a formal procedure for finding the best M, the method used was to simply
consider various possible Ms and see how well each matched the set of tokens which include
S. This time-intensive process was simplified somewhat by homemade tools to help find and
quickly listen to all tokens sharing some phonetic property or semantic annotation. The
17