ATTRIBUTION OF MUTUAL UNDERSTANDING 393
Dialogues are reordered by generating new start-times and
durations for each utterance as turn indices. The times and
durations are selected using random generators based on
parameters that depend on the values in the original
conversation. Thus, for each utterance ui a re-indexing úi is
constructed. The ú are sorted on their temporal indices. Analysis
of duration of overlaps is enabled by this framework but not
conducted here given that the transcripts addressed are not
annotated temporally beyond the relative order of turns. Where
temporal annotations are available, the analysis of
synchronization may have greater depth with the inclusion of
consideration of temporal overlap. In the reordered dialogue,
counts of allo-shared tokens and self-shared tokens are recorded,
just as with actual dialogue. The variables measured and
analyzed here are as specified in Table 2. The results allow for
the depiction of many contrasting proportions; however, the
specific contrasts of interest are whether actual repetition of
unigrams and n-grams for larger values of n exceeds the random
counterparts for any speaker. Thus, the null hypotheses tested in
each dialogue are as in (5) and (6).
(5) Randomized.Speaker.1 – Actual.Speaker.1 ≥ 0
(6) Randomized.Speaker.2+ – Actual.Speaker.2+ ≥ 0
The data is analyzed in each case using a generalized linear
model with a binomial error family.^36 Adjustments are made for
multiple comparisons using directed tests for significance,
wherein the null hypothesis essentially is that where
DialogType = Randomized repetition will equal or exceed
repetition for the corresponding Actual case.^37
(^36) Within R, this is using the following:
glm(OSprop~DialogTypeSpeakerNbar,family=binomial) and
glm(SSprop~DialogTypeSpeakerNbar,family=binomial).
(^37) With the R multcomp package, the following representative constructs
are used:
fos <- interaction(DialogType,Speaker,Nbar),
mfos <- glm(OSprop~fos,family=binomial),
mfos.mc <- glht(mfos,linfct = mcp(fos = "Tukey"),alternative="l").
See FRANK BRETZ ET AL., MULTIPLE COMPARISONS USING R (2011).
Subsequently, all tests are discarded which do not hold constant Speaker and