Transcript exploring reality tv data - AAAC emotion

EXPLORING REALITY TV DATA
Roddy Cowie & Cate Cox
Why reality data? How real does it need to be?
We want to look at emotion in action and interaction
Not episodes contrived to produce ‘pure’ emotion,
because most of the time, emotion shares a person’s
head with other constraints (plans, demands, etc).
There is no guarantee we will recognise emotion as it
appears in ‘headshare’ mode by collecting
samples of
pure emotion
Unfortunately, we don’t know very much about emotion
as it appears when people are engaged in ‘natural’
action & interaction.
So, this end of the project is about situations where
emotion appears in action and interaction.
How we have begun to look at reality TV data
What we have been looking at:
Approximately 5 hours of pre-edit footage from a television series
called Castaway 2000
The TYPE of data can be described as complex, within the terms of the
exemplar, i.e.,
Definitely provocative – forces one to expand and restructure
thinking about data,
Definitely falls within the WP5 remit of emotional behaviour in the
type of contexts which normally surround emotion in everyday life,
i.e., in interaction with other people and in the activities which we
conduct in a normal day
Without attempting to describe or label the data/footage at this
stage, in any particular way, purely through prima facie
observation the data/footage falls into 5 broad categories so
far:
1. Personal individual interviews with candidates, on their own,
being interviewed by a production crew member
Demo1.mpg
2. Interviews after having been through a fairly traumatic
challenge
Demo2.mpg
3. Group interview footage, where the group is addressed, where
an individual is focused on the context of the group, where an
individual is removed slightly away fro the group
Demo3.mpg
4 Group interaction
5. Field challenges
Demo3a.mpg
Demo4.mpg
Demo5.mpg
Demo5a.mpg
How we have begun to describe/approach the data
Basic approach at this stage:
let the footage ‘speak for itself’ - suggest categories/classification
BUT we need to set up ‘filters’ so that the messages are possible to
handle
‘just listening’ pulls you all over the place.
(Hansen – all data are theory-laden – the trick is to get a
decent match)
Develop a systematic approach without loading it too heavily with
preconceived notions of verbal categorical labels.
In mind of the exemplar - developing new labelling techniques in tandem
with new data, without yielding too many labels or developing in an ad-hoc
or un-coordinated way.
Trying to ‘unpack’ the process a little, initially by adapting the dimensional
approach in tandem with a coarse to fine approach
– getting coarse description allows you to get to know the data
and to focus in a systematic way on the informative parts.
A broad dimensional approach
As a first step, we are looking at the data without trying to classify
or describe an emotion, but simply to establish whether there is any
emotionality there at all.
We start with an adaptation of Feeltrace which presents one
dimension only: E-Trace
Definitely Absent vs Definitely Present
There are some subtleties in here – getting a scale which is intuitive
to use.
Definitely
Absent
Definitely
Present
Definitely
Absent
Definitely
Present
not edge towards edge towards
yes
not
What kind of information does this give us?
low incidence of unequivocal emotionality
clear absence of emotionality is equally rare
we see where to look for emotionality (eg long section at the end)
and how it is distributed (different rhythms in different episodes)
R
1.0
0.5
0.0
1
1415 2829 4243 5657 7071 8485 9899 11313 12727 14141 15555 16969 18383 19797 21211 22625 24039 25453 26867
Fig 1. Graph Representation E-Trace Tape 1_1
0.3
0.2
0.1
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Fig 2. Graph Frequency Distribution E-Trace Tape 1_1
Upper limit of confidence category
1
t
The exercise has already thrown up some other dimensions that are
critical:
Is ‘it’ discernable
Contextual Issues
yes/no which is a different question to
whether emotionality is there or not,
you may feel that it’s there but it’s
not clearly discernable
Coarse to fine context coding starts to emerge
Group/non-group
yes/no group dynamics come into play
which seems to lead the interaction
into a different area.
Other Dimensions being considered as they have emerged (so far)
from the data, whether emotionality is:
overt vs masked
mixed vs clear
genuine vs fake
……….
More Questions…
There are still many questions that will arise as we continue to explore
this type of data:
What to do with the data?
Physically separate clips (BT database model)? – or
Data Mining & Extraction mechanism (Like Ferret)?
Bearing in mind that this kind of naturalistic data shows emotional
behaviour that may not be discrete, constant or of short duration.
So, we are exploring how to explore.