jAL (print) issn 1479–7887

Journal of jAL (online) issn 1743–1743

Applied
Linguistics

Article

Metaphor clusters in discourse

Lynne J. Cameron and Juurd H. Stelma

Abstract
The phenomenon of clustering, where speakers or writers suddenly produce multiple
metaphors, is widespread and intriguing. This paper presents an innovative visualisa-
tion methodology for identifying and exploring metaphor clusters, comparing it to
existing methods that use cumulative frequency graphs and Poisson curve fitting, and
addressing issues that arise from these. Identification of clusters from the visualisation
is shown to be reliable and practical, while also offering in-depth exploration across
a range of discourse parameters.
Conversations aimed at conciliation between a perpetrator of violence and a vic-
tim (total 160 minutes) are analysed for clusters and their discourse functions. All
techniques show clusters at two distinct time scales, of around one minute and of
several seconds. Clusters in conciliation talk account for about 42 per cent of the total
metaphors, and cover about 30 per cent of the discourse. Discourse work carried out
in clusters includes explanation of a speaker’s perspective to the Other, appropriation
of metaphors originally used by the Other, and exploration of alternative, negative,
scenarios that had been possible choices for the speaker but had been rejected.
The finding that metaphor clusters are sites of intensive work relating to the central dis-
course purpose supports cluster exploration as a heuristic tool for discourse analysis.
Keywords: metaphor; conciliation; clusters; bursts; visualisation; discourse

Affiliations
Corresponding author: Professor L. J. Cameron, School of Education, University of Leeds, Leeds LS2 9JT, UK.
email: [email protected]
Juurd Stelma: School of Education, University of Manchester, UK.

JAL vol 1.2 2004: 107–136

©2004, equinox publishing
108 Metaphor clusters in discourse

1 Introduction
The methodological concern of this paper has arisen in the process of research-
ing metaphor in discourse. Several researchers have noted that metaphors are
not evenly distributed across talk or text, but come in clusters or bursts at certain
points, and may be nearly totally absent at other points (Cameron, 2003; Corts
& Meyers, 2002; Corts & Pollio, 1999; Low, 1997). We also know from empirical
work that people not only use metaphor in both poetic and prosaic discourse,
but that the conventionalisation of metaphor into language makes it almost
impossible to avoid metaphor. Contemporary metaphor theory (Lakoff, 1993)
attributes the ubiquity of metaphor in language use to the human need to con-
ceptualise the world through metaphor. Ubiquity combined with uneven spread
suggests something interesting may be happening. Why do people suddenly
produce clusters of metaphors at certain points in their speaking or writing?
What interactional or discourse work is being done in and with these clusters?
Is the clustering phenomenon similar in different types of discourse?
The phenomenon of metaphor clustering has been found in different dis-
course contexts, and in writing as well as in talk (Low, 1997; Cameron & Low,
2004; Koller, 2003). Clustering seems to have the potential to reveal something
of the conceptualisation and thinking processes of speaker or writer, at points
in talk or text where producers do something ‘out of the ordinary’ with meta-
phor. However, investigation of clustering phenomena first requires reliable
identification. Basically, we need to know how to distinguish a ‘cluster’ from its
surrounding talk or text. That is, we need information about the distribution
of metaphors across the talk or text in order to decide when the density of
metaphor changes sufficiently to warrant being labelled a ‘cluster’ or ‘burst’.
This paper takes forward the identification of metaphor clusters in discourse
by introducing a new computer-generated visual display technique that enables
researchers to find and investigate metaphor clusters. The visual display method
is shown to offer ‘reliable enough’ identification of clusters when compared with
other available methods, with the additional advantage of facilitating deeper
exploration of how metaphors are used through information about discourse
parameters and metaphors that remains available within the visualisation. In
addition, the paper contributes to our understanding of how clusters operate in
the dynamics of discourse by investigating a new discourse type: talk towards
conciliation between a perpetrator of violence and a victim.
A recent research project1 has provided us with the opportunity to develop
the visual method for identifying metaphor clusters in discourse and to explore
their occurrence in the context of conciliation conversations between a former
member of the Irish Republican Army, Pat Magee, and the daughter of one of
the victims of the bomb he planted, Jo Berry. More details of the context and
 L. J. Cameron & J. H. Stelma 109

data will be provided in a later section, but it may be useful for the reader to
see an extract from the conciliation talk here. Extract 1 comes from early in
their second conversation and includes a metaphor cluster which begins during
a turn of Jo (J) and continues across two changes of speaker, to the bomber
(P) at line 110 and back to the victim at line 127. Metaphorically used words
and phrases are underlined in the extract; more details on the identification of
metaphor are given in a later section.

Extract 1: From conciliation talk

87 Jo ...(2.0) but as I said to you,

88 s- something I’ve wanted,
89 ... almost since the moment the bomb went off.
90 ..I knew,
91 ...(2.0) back in the moment,
92 wh- what I wanted to do,
93 ... was bring as much –
94 ...(2.0) something –
95 .. as much positive out of it as I could.
96 .. you know,
97 Pat [hmh]
98 Jo ...(1.0) [and] I –
99 and I saw very clearly.
100 ...(1.0) that the –
101 .. the end of that journey,
102 would be,
103 .. sitting down and,
104 ... talking to the people who did it.
105 Pat .. hmh
106 Jo .. that just came in a moment,
107 and then went away,
108 and then –
109 .. there’s been a long long .. 16 years of [getting to this point].
110 Pat [hmh hmh]
111 ...(2.0) I suppose there’s a question,
112 that er comes to my mind from that.
113 .. er I don’t ...(1.0) think we’ve .. covered before.
114 ...(1.0) um er –
115 ... did you see it as like individuals,
116 or did you see it as a sort of a –
117 ...(1.0) the big .. political picture,
118 the IRA,
119 or,
120 .. the war.
121 ...(1.0) um you know what I mean,
122 er,
110 Metaphor clusters in discourse

123 Jo yeah,
124 Pat ... you were –
125 you were aware that there’s a –
126 .. it’s going to be an individual who you’d be sitting down with.
127 Jo hmh
128 ...(1.0) I saw it as both.

Clustering of metaphors seems to occur on several scales of talk. Inside the level
of the episode in Extract 1, which is about one and a half minutes of talk from a
50-minute conversation, we find smaller clusters within one or two utterances.
An example can be seen in lines 115–7, where see is used metaphorically to talk
about understanding and picture to talk about an historico-political context. In
a study of metaphor in classroom interaction, Cameron (2003) found that, on
this micro-level scale, two or more metaphors were more likely to occur than
single isolated metaphors. Moreover, metaphors that occurred ‘in company’
were, on average, found in groups of 3.5, indicating that co-occurrence was
not a gradual phenomenon but a distinct one.
Our theoretical framework for the investigation of conciliation talk holds that
discourse can be seen as a complex system, which is dynamic, non-linear and
adaptive (Cameron, 2003). Within a complex system, self-organisation and the
emergence of new forms or paths (rather than linear cause and effect) is the
process that links different levels or scales. From a complex systems perspec-
tive, metaphor clusters are emergent outcomes of the system – that is the two
speakers in interaction, adapting and self-organising in response to interactional
demands. Complex systems typically move between variability and stability, and
metaphor clusters are likely to be sites of particular variability or flexibility in
the interactional process that involves speakers in both talking and thinking.
To understand clustering, we need to explore the discourse activity that takes
place inside them and to trace the interactional demands that motivate them.
Researching discourse as a complex system suggests methodological principles
and constraints, including: attention to dynamics, processes and change; the
inseparability of data and context; the analysis and exploration of data as it is with
the ‘noise’ and mess of real talk, rather than ‘cleaned up’. The use of visualisation
as a tool for investigating discourse dynamics is partly motivated by visual modes
of representing complex systems in other disciplines, such as biology.
The paper begins with a review of published studies reporting metaphor
clustering, and of their methodologies for identifying clusters. From this review,
a number of key issues emerge that any method needs to address. We then
proceed to explain our visual display technique and its use with data from
conciliation meetings. We show how clusters appear in the visual display, and
compare visual identification with graphical and statistical procedures. While
the aims of the paper are primarily methodological, the paper also reports on
 L. J. Cameron & J. H. Stelma 111

the nature of clustering in conciliation talk, a discourse context with its own
particular interactional demands. The findings lead to a broader view of what
motivates discourse participants to produce clusters of metaphors and what
they may achieve in doing so. We conclude by suggesting how researchers might
use metaphor clusters as a heuristic tool in discourse analysis.

2 Metaphor clusters in discourse: previous studies and

techniques for identification
The literature offers two main techniques for identifying clustering phenomena
in the use of metaphors: cumulative frequency graphs and a statistical pro-
cedure. In this section, we review published studies and their use of the two
methods, identifying methodological issues in need of attention and bringing
together findings on the nature of discourse inside clusters.
To introduce the concept of a cumulative frequency graph, Figure 1 shows
the production of metaphors in the course of a primary school lesson (from
Cameron, 2003: 123).

Figure 1: Cumulative frequency graph for classroom talk (adapted from Cameron,
2003: 123)

The horizontal axis shows time, here operationalised as number of words from
the beginning of the transcribed talk. At any point along the horizontal axis,
the vertical axis coordinate of the graph is the cumulative number of metaphors
112 Metaphor clusters in discourse

used in the talk from the beginning, up to and including this point. The graph
will continue upwards as more and more metaphors are used. If there is a patch
of talk without any metaphors, the graph will go flat and proceed horizontally.
A clustering of metaphors in the talk is shown by a sudden steep climb in the
curve. Identification of metaphor clusters can be done by visual inspection of
the cumulative frequency curve, judging where the gradient changes sufficiently
from the average slope. In this discourse event, seven steep increases are seen
on the graph, and these are boxed in Figure 1.
Inside these clusters, as in the other lessons investigated, metaphors are used
mainly to explain unfamiliar or difficult topics, and for discourse management
in the openings and closings of lessons.
The cumulative frequency approach was used by Pollio and Barlow (1975)
in a report on figurative language in a psychotherapy session. Taking a broad
category of ‘figurative language’ that includes metaphor, litotes, oxymoron and
metonymy, they plot ‘the total figurative output for the patient and therapist
over consecutive utterances’ (1975: 240) to produce cumulative frequency
graphs. Bursts, as they call clusters, in the patient’s use of novel figurative
language allowed the identification of major movements in interpretive activity.
Professionals who reviewed the talk considered that the bursts coincided with
periods of ‘substantial therapeutic insight’ (Corts & Meyers, 2002: 392).
Corts and Pollio (1999) report bursts of figurative language, as well as gestures,
in a study of three college lectures. This study introduces a statistical method
for identifying clustering of figures and gestures (‘backed up’ by cumulative
frequency graphs). The method involves identifying statistically significant
increases in the rate of metaphor use, rather than judging increases visually.
The nature of the distribution of figurative language across a discourse event
has to be established e.g. a normal, binomial, or Poisson distribution, and then
the fit of instances of actual use to the distribution is tested. A cluster or burst
is identified where the probability of an increase in frequency arising from
chance is less than the chosen significance level. The method is based on the use
of centred moving averages (CMA) of instances of figurative language across
successive intervals of the discourse:
A CMA is calculated by taking an arbitrary number of sentences, in this case
five, and calculating the mean number of figures in that group of sentences.
Once a value had been obtained for the first five sentences, it was next
calculated for Sentences 2 through 6, Sentences 3 through 7, and so on until
the means of all possible sets of five consecutive sentences in the corpus had
been computed. (Corts & Pollio, 1999: 86. Our emphasis)
The authors state that the use of figurative language and gestures by the lecturer
in their data followed a Poisson distribution, although they do not give details
 L. J. Cameron & J. H. Stelma 113

of how this was established. Having calculated the probabilities that specific
values of CMA could occur by chance from a Poisson distribution, the authors
took probabilities of 0.10 or less as indicating the presence of a burst in the
output of figures and gestures.
We need to consider the appropriacy of the Poisson distribution to describe
the occurrence of metaphors in discourse. The equation was developed origi-
nally to describe the phenomenon of Prussian army officers being kicked by
their horses. It describes a situation in which there is a very low probability of
an occurrence of a phenomenon across a high number of opportunities. From
this perspective, the occurrence of metaphors in a discourse event would seem
to match the conditions for a Poisson distribution. There are, however, other
conditions on a Poisson distribution that make it appear less appropriate. Events
that can be described by a Poisson curve occur at random, singly, uniformly
and independently. The use of a metaphor may be random (mathematically,
rather than discoursally), but it is seldom a single event in spontaneous talk;
is often not independent – as when a series of connected metaphors is used
to talk about a topic; and the existence of clustering denies uniformity of
occurrence. While acknowledging doubts about the appropriacy of the Poisson
distribution to describe the occurrence of metaphors in discourse, we explored
further how the statistical procedure might contribute to the identification of
metaphor clusters.
When bursts of figurative language in the college lectures were examined
for content and purpose, they were found to be ‘topical’, i.e., connected to the
main topics of the lecture rather than to diversions, and to be particularly used
for topics that were more complex, unfamiliar or abstract for the students, e.g.
ageing or alcoholism. It was also found that, although the periods of bursts
only accounted for a small percentage of the total lecture, they included the
majority of ‘novel’ metaphors. Corts and Pollio conclude that ‘bursts of figura-
tive language are characterised by a coherent metaphor that is then elaborated
by an interrelated network of ideas and images’ (Corts & Pollio, 1999: 96). The
topical, novel and coherent nature of figurative language in bursts lead Corts
and Pollio to suggest that bursts of figures in lectures, as in school talk, primarily
assist understanding and recall of difficult or new content. They also found that
figures of speech were used in ‘structural mode’: ‘to organize and direct the
flow of the lecture’ (1999: 97), similar to the discourse management found in
openings in classroom lessons, but that, unlike the classroom discourse, these
figures did not tend to occur as densely in the lectures. This is probably because
the management task of college lecturers is usually less demanding than that
of school teachers.
A partial replication and extension study (Corts, 1999, reported in Corts &
Meyers, 2002), which examined how far bursts might be a feature of individual
114 Metaphor clusters in discourse

style or of the lecture topic, again found bursts to occur in explanations of a

single topic and to consist mainly of metaphors, similes and analogies (rather
than other figures) linked to a single root metaphor. This study found two
further types of bursts that were less coherent than the single conceptual type.
One of these combined more than just a single root metaphor, for instance in
comparing two possible metaphorical conceptualisations of a topic. Corts and
Meyers (2002) comment, ‘in some of these bursts, there are actually two smaller
bursts…with very few sentences between’ (2002: 394). This comment raises
the issue of where the boundaries of bursts or clusters are placed and whether
all sentences included in a burst are required to contain metaphors. The other
type of burst occurred when a speaker summarised multiple metaphors at the
beginning or end of a lecture.
Corts and Meyers (2002) investigated metaphor clustering in a further dis-
course context, that of Baptist sermons, using visual inspection of cumulative
frequency graphs2. The statistical method was dropped in this study for two
reasons. Firstly, in discourse with a high level of figurative language throughout,
Corts and Meyers say that there would be a ceiling effect that might prevent the
identification of important bursts. In fact, a better argument might be that the
higher density of metaphors in the discourse rendered the Poisson distribution
inappropriate. Secondly, they argue that the moving average technique identi-
fies changes in production rate, regardless of their nature, and so some bursts
may be the result of ‘random variation in the quantity of figurative language’
(Corts & Meyers, 2002: 394). They therefore used cumulative frequency graphs
of figures of speech, rated for topicality, coherence around a root metaphor, and
novelty. This decision would not be acceptable for a researcher working within a
complex systems framework, and we address this further in the next section.
Clusters (rather than ‘bursts’, which they restrict to statistically-identified
increases) were identified by inspecting the topical stream of the output curves,
i.e. the construct of clusters was partly defined by topicality in this study. Two
independent raters then judged which sentences in the sermons corresponded
to the beginnings and endings of clusters from the cumulative frequency graphs.
Clusters were found to occur across the sermons, with between two and four
times the number of figures per sentence inside clusters as in non-cluster talk.
Finding clusters in sermons as well as in college lectures and school classroom
talk adds to the evidence that clusters occur in different discourse types. All
three discourse types, however, share the dimension of a single main speaker
and low levels of interaction.
Inside sermon clusters, as in college lectures, there was a tendency for figura-
tive language to be topically coherent. Novelty, however, was no longer a key
factor, with novel metaphors found outside clusters as well as inside. Clusters
often arose when a conceptual, or root, metaphor was needed to talk about
 L. J. Cameron & J. H. Stelma 115

something e.g. a payment metaphor to talk about payment to God. Corts and
Meyers note that neither topicality nor coherence alone could account for the
increase in figurative language in clusters, but rather both acted together in the
representation of a topic through a single conceptual metaphor. They suggest
that the fact that the speakers in both lectures and sermons are ‘experts’ in their
field talking to less expert listeners may lead to multiple instances of explana-
tions, which in turn produce these topically coherent clusters. Some clusters
were not explanatory, but had some other function, including persuasion
through rhetorical repetition. In addition, they note that most clusters reflect
changes in the purposes or intent of the speaker, which would seem to imply
that some discourse management was involved.

3 Issues raised by previous studies

Pulling together previously published studies gives us a picture of metaphor
clustering in discourse, while raising further issues, particularly methodological
ones. In this section, we outline these points, before proceeding to report how
they were addressed in the conciliation project.

3.1 The discourse work of metaphor clusters

A somewhat coherent picture of metaphor clustering emerges from the studies
of classroom lessons, college lectures, psychotherapy interviews, and Baptist
sermons. Each type of discourse contains clusters or bursts of metaphors,
suggesting that clustering is a general phenomenon that will be found across
discourse types. Further, it seems that clusters of figurative language, often
metaphorical, play an important role in the unfolding of the discourse – or,
in complex systems terms, in the trajectory of the interactional system. For
example, clusters in lectures, lessons and sermons often centred around the use
of one or more conceptual metaphors, extended in order for the more expert
speaker to explain various aspects of the topic to the listeners. Rhetorical moves
such as summarising, parallelism and contrast increased the use of topically
coherent metaphors at certain points in the discourse. Explanation of difficult
or unfamiliar ideas – the main discourse work carried out in clusters – might
be seen also as the central purpose of lectures, lessons and sermons. In the
psychotherapy interview too, the generation of important therapeutic insights in
clusters of figurative language pushes forward the central discourse purpose.
The evidence also suggests that discourse management, or ‘structuring’, gives
rise to increased metaphor use. The greater organising demands of the school
classroom seems to have led to more clusters linked to discourse management,
in comparison with the college lectures and sermons.
116 Metaphor clusters in discourse

However, metaphor clusters have so far only been explored in a limited

range of types of spoken discourse. The conciliation talk offers the opportunity
to investigate clustering in a new type of discourse, with different purposes,
different levels of formality and much more speaker interaction. In contrast
with the expert-novice scenarios of sermons, lectures and classrooms, the
two speakers involved in the conciliation talk are equally ‘expert’ in what they
talk mostly about – their personal histories and experiences. In contrast with
the content of the other studies, the topics that they talk about are painful,
complicated and personal. The broad purpose of the discourse is not primarily
to transmit information or persuade, but is to reduce alterity, or Otherness,
between participants (Cameron, 2003; Wertsch, 1998); in the complex dynamic
system of the conciliation talk, participants begin from very different starting
points and want to move towards greater understanding of ‘the Other’. The use
of data from three discourse events separated by months and years also allows
investigation of the discourse dynamics on several timescales.

3.2 What counts as metaphor

Decisions as to what counts as metaphor or as figurative language are basic
to the investigation of clusters. Comparability across studies require the same
or equivalent units to be used. If studies define ‘metaphor’ differently, then
generalising across them to claims about the use of metaphor in discourse
becomes tricky. The central issue seems to be the breadth of the categories of
‘figurative language’ and ‘metaphor’.
The studies of Pollio and Barlow (1975) and Corts and Pollio (1999) started
from a broad category of figurative language, and found that conceptual or
root metaphors, extended as metaphors, similes and analogies, accounted
for the majority of figurative language in clusters. Corts and Meyers (2002)
made a methodological decision to start from this sub-category of figurative
language, root metaphors linked by shared content, and to find clusters using
the cumulative frequency graphs of the restricted sub-category. The category
was restricted early in the research process in order to avoid the inclusion of
clusters generated by ‘random variation’. However, over a collection of studies,
what looks like random noise in data from one or two studies may begin to
look more interesting. Because of this, a principle of researching complex
systems is that data should not be discarded too early (Thelen & Smith, 1994).
By taking topicality as a defining feature of clusters, Corts and Meyers’ findings
are somewhat pre-determined and what might be interesting data has been
discarded. We suggest that decisions about the quality of bursts or clusters
should be made at a later stage of analysis, rather than being built into the
identification stage.
 L. J. Cameron & J. H. Stelma 117

A broad category, which can be sub-divided in various ways, seems to be

the most useful to work with. In the conciliation project, we restricted our
concern to a broad category ‘metaphor’, excluding non-metaphorical figurative
language, such as oxymoron. ‘Metaphor’ includes various surface language
forms such as similes, analogies and hyperbole, but only those instances which
are metaphorical, in the sense of bringing two disjunctive domains together.
This category would appear from the available information to be roughly
equivalent to the metaphors, similes and analogies that the studies reported
above found to be constituting bursts of ‘figurative language’, and much broader
than the subset of root metaphors and extensions used in the Corts and Meyers’
frequency graphs.

3.3 The use of statistics

The Poisson technique appears to offer a way to increase reliability by providing
a statistical test of the size of increases in frequency of metaphor use. However,
the report of the technique in Corts and Pollio (1999) raised further issues. The
paper did not give precise details of how to carry out the procedure, so these
had to be worked out. Further, we had some doubts as to whether or not use of
metaphor in talk does actually follow a Poisson distribution, particularly when
a broad category is used, so we needed to find out how to test this. We also saw a
need for a more motivated method for deciding on the size of intervals to average
or sum, rather than the arbitrary use of consecutive intervals of five sentences.

3.4 Units of discourse

The identification of metaphor clusters or bursts requires the analyst to decide
which increases in use are sufficiently sudden or distinct from the immediately
previous discourse to warrant labelling as clusters. By plotting the cumulative
frequencies on a graph, increases in use appear as sudden rises in the curve,
i.e., they become visually identifiable. However, reliability and validity in visual
identification are affected by the unit chosen to divide the discourse.
The studies made different decisions about the unit of discourse on the
horizontal axis of the cumulative frequency graph. Variation in the length (in
words) of the discourse unit, across the event under investigation, can have an
impact on the interpretation of the graph. The psychotherapy study (Pollio &
Barlow, 1975) used consecutive ‘communications’, or ‘utterances’, as the unit
of talk between the therapist and the patient. Utterances were bounded by
a change in speaker, regardless of the length of time a speaker was talking.
Their unit of analysis, therefore, is similar to what would currently be called a
turn. Pollio and Barlow (1975) report that the average turn/utterance length
118 Metaphor clusters in discourse

for the patient in their data is 30.5 words, with about 24 words per turn in the
beginning of the meeting to just above 48 words per turn in the middle part,
and then about 38 words per turn towards the end of the meeting. Pollio and
Barlow also report numbers for novel figures per word, which appear to vary in
a similar way across the three stages of the meeting. However, what they graph
is the number of novel figures per turn/utterance. The changes in the slope of
their cumulative frequency graph therefore reflect the variation in their unit
of discourse as well as variation in output of novel figures.
The later studies by Corts and Pollio (1999) and Corts and Meyers (2002)
take the (spoken) ‘sentence’ as their unit of discourse. Since a turn may be sub-
divided into more than one sentence, this seems an improvement on the earlier
use of turns as a unit of discourse. It may also be reasonably appropriate, since
lectures and sermons are formal and probably prepared in writing. However,
sentences, particularly those produced by expert presenters, are known to use
variation in length as a rhetorical device. In graphing the number of figures
per sentence, we still have a varying unit of discourse, rendering comparison
across cases impossible. If we are dealing with less formal spoken discourse than
lectures or sermons, the ‘sentence’ becomes even less appropriate as a unit of
analysis. In an attempt to move to a more universal unit and aiming for some
comparability of scale across discourse contexts, Cameron (2003) took a unit
of 10 transcribed words and graphed these in blocks of 5, i.e. 50 spoken words.
While being roughly equivalent to the 5 sentence block used by Corts and Pollio,
this unit helped to even out variation in length across the discourse event. The
average speaking time for 50 words, including pauses, was 35 seconds.
Comparability across studies of different discourse contexts requires a unit
of discourse that remains constant and fixed within and across those contexts.
Utterances, turns, and sentences all vary within the discourse context. A fixed
number of transcribed words is less variable. A time measure, such as seconds,
would remove variability altogether, but would not be sensitive to the nature of
the discourse. In deciding on a unit to use for the data in the conciliation project,
we needed to decide how to balance comparability with discourse sensitivity.

4 Identifying metaphor clusters in the conciliation talk

using a visual display
We came to the conciliation data with the above methodological issues needing
to be resolved. We wanted to identify metaphor clusters in the talk because we
believed that important discourse work was taking place within them, but we
first needed to find practical, reliable and valid ways to identify them. We now
report how we went about this, and how the visual display methodology offers
an alternative to frequency graphs and statistical procedures.
 L. J. Cameron & J. H. Stelma 119

4.1 The participants and data

The participants in the conciliation talk were Jo Berry and Pat Magee, a former
member of the Irish Republican Army who, in 1984, planted a bomb which
killed several members of the British government including Jo Berry’s father.
After early release from prison in 1998, Pat Magee agreed to meet Jo Berry to
talk about his motivation for the bombing. They have since participated in
a series of meetings, some of which were filmed for a TV documentary, and
have given several joint interviews and talks. Jo Berry provided us with video
recordings of two conversations (50 minutes and 110 minutes) that took place
in late 2000 and early 2001. These conversations were recorded during the
second and fourth meetings between Jo and Pat. The transcribed conversations
comprise the data used in this article.3

4.2 Transcription and choice of unit of discourse

A first task in preparing to explore metaphor clustering was to choose a unit of
discourse that would be consistent in length of time across the talk, and thus
minimise the effect of variation in length on identification and interpretation
of clusters. Our choice of unit for transcribing the spoken conciliation data was
the intonation unit in the form developed by Chafe (1994). Intonation units
have a psycholinguistic rationale for use in research on spoken interaction,
and this rationale can be drawn upon in interpreting any metaphor clustering
phenomena that are identified. As talk proceeds, ideas move in and out of a
speaker’s conscious attention. Chafe suggests that ‘an intonation unit verbalises
the speaker’s focus of consciousness at that moment’ (1994: 63). Each line
in Extract 1 is an intonation unit (IU)4. Identification of IUs uses various
prosodic criteria. Prototypical intonation units often begin with a change in
tempo (anacrusis) and/or a resetting of the pitch baseline, and usually have an
identifiable terminal pitch contour (falling or rising). They are often preceded
or followed by a pause (but may also contain pausing within their bounda-
ries) and sometimes begin and end with a creaky voice, which may obscure
terminal pitch contours (Chafe, 1994: 53–62). This multi-dimensional nature
of intonation units means that, unlike pauses for example, they cannot always
be identified objectively or mechanically. However, with training, high levels
of intra-rater reliability, and reasonable levels of inter-rater reliability appear
possible (Stelma & Cameron, in preparation).
Our decision to use the intonation unit was made on the basis that it is both
a unit in which the cognitive and the verbal overlap, and a good approximation
to an absolute measure that will enable valid comparisons across different
discourse contexts. Because the intonation unit is a consistently small unit of
120 Metaphor clusters in discourse

spoken discourse (around 1–2 seconds), its use allowed more detail in our iden-
tification of clustering phenomena, in contrast to sentences or turns. Moreover,
the fact that intonation units are relatively consistent in length, in contrast with
turns in particular, ensured that the cumulative frequency graph would reliably
reveal clusters, and not changes in the length of the unit of discourse.
Training in the identification of intonation units and double checking of
portions of the transcriptions were carried out to increase reliability. The 50
minutes of the first conversation was transcribed into 2769 IUs. The second
conversation, which was 110 minutes long, was transcribed into 4043 IUs.

4.3 Identification of metaphor

Once the data had been transcribed, we identified linguistic metaphors in the
data following the method set out in Cameron (2003). Words and expressions
in the talk, which can be seen as bringing some meaning from another, dis-
junctive and non-metaphorical domain, were underlined in the transcription
as metaphor ‘Vehicle’ terms. So in Extract 1, for example, pain is classed as
metaphorical since there is another distinct meaning – of physical pain – that
is brought to the topic of grief.
The identification of metaphors requires the analyst to make many decisions
about what is included and excluded. Our approach is informed by cognitive
metaphor theory, which holds that metaphor in language reflects metaphorically
structured patterns of thought. Metaphorical patterns of thought may or may not
be activated in the mind of an individual speaker engaged in interaction but they
retain the potential to become so (Gibbs, 1994). Thus we include as metaphorical
something when used to refer to a non-concrete entity such as a wish/desire
(Extract 1, line 88) and some prepositions, e.g. out of when used to refer to a
process that does not involve physical exiting of a container (Extract 1, line 95).
Some delexicalised verbs are included as metaphorical when used for non-physi-
cal, non-concrete processes e.g. came, went away to refer to an idea (Extract 1,
lines 106–7). Many of these words and phrases would be traditionally labelled
‘dead’ metaphors, but our concern is with metaphorical potential and not to
decide whether actual instances of production and interpretation involved active
metaphorical processing by participants. In fact, discourse data only seldom
offers evidence that would allow the analyst to infer active metaphorising.
The number of metaphors identified in the two conversations were 775 and
1384 respectively, with metaphor densities of 98 and 107 per 1000 words. As a
comparison, classroom talk in which metaphors were identified in a comparable
way produced a metaphor density of just 27 per 1000 words (Cameron, 2003).
By starting from a broad category of metaphor, we can later discard cer-
tain categories or analyse them separately. In our analysis, we moved from
 L. J. Cameron & J. H. Stelma 121

identification to the categorisation of metaphors in terms of their Vehicles.

Again we drew on cognitive metaphor theory in the construction of Vehicle
domains as we worked both inductively, grouping similar Vehicles together,
and deductively, using conceptual domains found in the literature, such as
journey. The process of grouping Vehicle terms and then Topic domains to
produce larger ‘systematic’ metaphors is unavoidably interpretive. Furthermore,
there is as yet no standard method for doing this. It would be possible to work
entirely inductively, allowing Vehicle domains to emerge from the data, or to
work more deductively, looking for Vehicles that fit domains described in the
cognitive metaphor literature. Our method is a compromise between these
extremes, guided by a discourse (rather than cognitive linguistic) principle
that Vehicle and Topic domains should stay as close as possible to the words
used by participants. Thus, the following Vehicles from Extract 1 were included
in a journey domain: the end of the journey, bring out of, long, getting to this
point, comes to; we did not however generalise further to domains such as
source–path–goal or space (Lakoff, 1993). The Vehicle terms saw, clearly,
see, picture were included in a domain labelled seeing.
Once Vehicle domains were identified, metaphors within them were grouped
by Topic, producing systematic metaphors5, such as understanding the
other is a journey; the process of conciliation is a journey; the on-
going talk is a journey.

5 Visualisation of the conciliation talk

The VisDis software6 creates a visualisation of the dynamics of metaphor in
discourse as it evolves over time. Shifts in the talk appear as shifts in the path
of the conversation on the computer screen or page. The software is based on
a spreadsheet facility which automatically counts and sums metaphors that
have been identified by underlining in the transcript. The researcher can assign
values to further variables, representing other aspects of the discourse, which
can be used to construct visual displays. Aspects of talk can be drawn on sepa-
rate displays or placed together on the same visual display to explore patterns of
co-occurrence. We coded and created displays for the following aspects of the
conciliation talk: instances of linguistic metaphor, Vehicle domain, discourse
topic and change of topic, speaker, and alterity (operationalised here as nominal
or pronominal reference to self, Other and others). Figure 2 shows a complete
visual display of a small part of the first conversation. Time, measured in IUs,
moves vertically from top to bottom in these displays. The talk of the speak-
ers can be seen moving through the alterity space between them and across
changes of topic. Metaphors are shown as diamonds on the left hand side of
the diagram.
122 Metaphor clusters in discourse

Figure 2: Visual display of a short section of conciliation conversation

J S A F O

Metaphor Vehicle domains

J = JOURNEY
S = SEEING
C = CONNECTION
A = PHYSICAL ACTION
F = FICTION / PLAY
O = other domains

Figure 3a: Visual display of metaphors in Figure 3b: Visual display of metaphors
first conciliation conversation: first 1000 in first 250 intonation units, sorted by
intonation units and first 250 intonation metaphor Vehicle domain
units
 L. J. Cameron & J. H. Stelma 123

Figure 3 shows the displays of metaphors produced for the first 1000 IUs (about
20 minutes) and 250 IUs of talk. Each diamond on the time line represents a
use of metaphor. Figure 3a shows the whole 1000 IUs in the left hand line, with
the second line magnifying the first 250 IUs. The stretch of talk in Extract 1 is
enclosed in dotted lines.
In Figure 3b, the diamonds in the right hand column are organised by Vehicle
domain (only the major domains are included here); in the computer display,
each domain is given a different colour. From left to right, these domains are
journey, seeing, connection/separation, physical action, fiction/play,
other. The software can automatically produce displays like Figure 3b once
metaphor domains are coded as a variable in the spreadsheet interface. It is
at this point in the investigation, in line with the principle of not jettisoning
data as ‘noise’ too early in the research process, that the researcher can decide
certain domains are more relevant than others, and focus on a subset of the
metaphors.
The visual display in Figure 3a offers a different way to see metaphor cluster-
ing from a cumulative frequency graph. Apart from the superficial difference in
orientation, which can be easily switched from vertical to horizontal, clusters
appear as dense collections of diamonds (or points, triangles etc) in the time
line of production, rather than as increases in slope. In Figure 3a, the major
cluster at around IUs 600–700 stands out as the most visible, with other clearly
visible clusters around IUs 100 and 400. At IU 587 in the conversation, Jo begins
to read aloud a poem about ‘building bridges’ that she had written in response
to her father’s death. The poem is densely packed with metaphors, hence the
cluster. Zooming in to the magnified 250 IU block clearly shows the cluster
that is Extract 1, between 87 and 120 IUs. Other possible clusters can be seen
around 40, 150 and 200 IUs.
Unlike a cumulative frequency graph, the visual display offers extra detail
on the domain content of clusters and how one domain links to another in the
dynamics of talk. For example, we found that the domain of seeing very often
co-occurs with the domain of journey. The software enables the researcher to
bring the relevant portion of the transcript up on screen alongside the visual
display; moving between visual and textual information showed that, across
the discourse event, there are three distinct ways in which participants connect
journeys and seeing, rather than the single connection put forward by cogni-
tive metaphor theory (that the two domains are part of a larger metaphorical
scenario in which travellers on a journey look ahead to see which way to go). The
visual display cannot replace analysis of transcribed talk, but it facilitates a much
richer exploration of discourse in moving between visual and transcription.
124 Metaphor clusters in discourse

The possibility of including a range of aspects of discourse in the visual

display allows other patterns and research questions to be investigated, such as
‘Do clusters tend to occur around changes of discourse topic and speaker?’
Metaphor clusters can be identified from the visual display of metaphors as
points on a time line. The facility of retaining rich information about a range
of metaphor and discourse parameters ‘underneath’ the visual display enables
researchers to go beyond identification into exploration.

6 Comparison of the visual display method of cluster identification

with other methods
In this section, we compare the use of the visual display with the two other
methods of identifying clusters found in the literature: cumulative frequency
graphs and the Poisson procedure.

6.1 Identifying clusters in conciliation talk with a cumulative frequency graph

The cumulative frequency graph of the entire 50 minutes of the first conversa-
tion between perpetrator and victim appears as Figure 4. On this scale, it is
difficult to see many changes in the gradient of the curve. Only between IUs
600 and 700 (the poem) does the graph have a clearly visible sudden increase
in slope (boxed as a cluster in Figure 4).
In order to reveal clustering on smaller timescales, we increased the ‘magni-
fication’ of the cumulative frequency graph. We experimented with resolutions
of 1000, 500 and 250 IUs. At a resolution of 250 IUs (comparable to Figure 3b),
a new level of clusters became visible. Figure 5 shows the metaphor use on this
timescale of the discourse, and representing the first 250 IUs of the discourse
event (i.e. about 5 minutes of spoken discourse), and including Extract 1. The
position of this segment in the total discourse event is shown by the dotted
outline in Figure 4. At this magnification we were able to identify four metaphor
clusters (marked A, B, C and D in Figure 5). Hence, exploring the discourse at
different magnifications confirmed earlier findings that clustering phenomena
occur on different timescales in discourse (Cameron, 2003). The detail and the
temporal accuracy that result from the use of intonation units in transcription
increase the validity of the observation that clustering phenomena occur on
different timescales.
As mentioned earlier, problems can arise with the boundaries of clusters. For
example, in Figure 5, clusters B and C are separated by about 18 IUs or 36 sec-
onds of talk and the researcher needs to decide whether they should be taken as
a single large cluster or two smaller ones with a gap between lines 118 and 138.
Returning to the visual display and looking at this point in Figure 3b, we can
 L. J. Cameron & J. H. Stelma 125

Figure 4: Cumulative frequency graph of first conciliation conversation (to end)

Figure 5: Cumulative frequency graph of initial 250 intonation units of first conciliation
conversation
126 Metaphor clusters in discourse

see a shift in Vehicle domains between the two clusters, suggesting that C is
indeed a separate cluster. Again, the extra information available from the visual
display makes it more useful than the cumulative frequency graph in which
such qualitative information has been lost in the process of quantification.
The visual display is perhaps more intuitive to less mathematically-minded
analysts than cumulative frequency graphs. Metaphor clusters appear as
‘absolute’ clusters in the visual display, whereas in the cumulative frequency
graphs they appear as changes in gradient and are thus seen relative to their
immediately preceding frequency level.
As we explored the data at different magnifications, it was often difficult to
determine whether an increase in the slope of the cumulative frequency graph
actually did indicate a cluster. Figure 6 represents the discourse data between
IUs 1000 and 1250 (i.e., the same magnification as in Figure 5). In this figure
there are definite changes in the slope of the cumulative frequency graph.
Potential clusters in this figure are marked E, F and G. However, whether E, F,
and G, or maybe only G, should be identified is difficult to decide.
While cumulative frequency graphs seem to offer a fairly straightforward
technique for identifying metaphor clusters, scale is clearly an important vari-
able, affecting what is noticed and what can be claimed as ‘visible’ to the eye
of the analyst. The problem with borderline cases motivated us to take a closer
look at statistical alternatives to identifying clusters in the data.

Figure 6: Possible clusters in segment of first conciliation conversation

 L. J. Cameron & J. H. Stelma 127

6.2 Identifying clusters in conciliation talk with the statistical procedure

In trying to replicate this methodology, we addressed problems that we found
with the Poisson procedure as reported by Corts and Pollio (1999). Intonation
units were used as a more absolute and comparable measure than sentences. In
computing Poisson probabilities (using Microsoft Excel spreadsheet software),
we found that moving sums, rather than averages were needed. Table 1 illustrates
our calculation of moving sums for a small segment of the data. For example,
the moving sum for the interval 93 through 97 is 2 because there are two meta-
phors in intonation units 93–97. In this table, no moving sums are calculated
for intervals before 91–95 because this data is not ‘visible’ here.

Intonation unit Metaphor use Interval Moving sum

91 1
92 0
93 1
94 0
95 1 91–95 3
96 0 92–96 2
97 0 93–97 2
98 0 94–98 1
99 2 95–99 3
100 0 96–100 2

Table 1: Calculation of consecutive moving sums

Before we moved on to calculate the Poisson probabilities, to determine whether

any of the moving sums occurred by chance or not, we investigated whether
the distribution of metaphor use was, in fact, following a Poisson distribution.
A Poisson distribution describes situations with a very low probability (p) of
occurrence of a particular phenomenon, and a large number of opportunities
for the phenomenon to happen, or cases (N). A Poisson distribution thus tends
to be skewed to the left when displayed as a graph. Having observed clustering
on different timescales when using the cumulative frequency graphs and visual
displays, we suspected that the extent to which the use of metaphor would
follow a Poisson distribution might depend on the size of the intervals specified.
Hence, we calculated moving sums for intervals of 5, 10, 15 and 20 IUs. Figure 7
illustrates how the distribution of metaphors in our data (the bars) compares
to a theoretical Poisson distribution (the curves) for the four different moving
sum intervals that we calculated.
We selected not to test intervals that were larger than 20 IUs. This was because
with such larger intervals there would be a much higher probability (p) of
having metaphor clusters in any one interval. In other words, the distribution
is moving away from a Poisson distribution towards a normal distribution. In
128 Metaphor clusters in discourse

Figure 7: Comparisons of distributions of metaphors in first conciliation conversation

Figure 7 one can see that as the interval increases, the distributions (both in
our data and in the theoretical Poisson distribution) are becoming less skewed
to the left, and more similar to a normal distribution.
To test whether metaphor use followed a Poisson distribution we calculated
the goodness-of-fit between the distribution of independent intervals of moving
sum values and a theoretical Poisson distribution. That is, for intervals of five
IUs we used the distribution of the moving sums for the intervals 1–5, 6–10,
11–15, and so on (Figure 7 also represents these independent intervals). The
theoretical Poisson distribution was generated mathematically using the mean
of the independent moving sums we wanted to test. We repeated this for each
of the interval sizes that we wished to test (5, 10, 15 and 20).
A chi-square test for the intervals of 5 and 20 IUs returned not significant
results: [χ2(4, N = 553) = 9.34, p>.05] and [χ2(8, N = 138) = 12.43, p>.05] for inter-
vals of 5 and 20 respectively. In other words, there is no evidence that metaphor
use measured in intervals of 5 or 20 IUs does not follow a Poisson distribution.
By contrast, the chi-square test for intervals of 10 and 15 IUs returned significant
results: [χ2(6, N = 276) = 16.52, p<.05] and [χ2(7, N = 138) = 12.43, p<.05] for
intervals of 10 and 15 respectively. Hence, metaphor use measured in intervals
of 10 and 15 IUs may on this basis be rejected as Poisson distributions. As is
 L. J. Cameron & J. H. Stelma 129

common practice in preparing contingency tables for the chi-square test, when
an expected frequency (generated by the theoretical Poisson distribution) fell
below 4, we combined this with the next larger value in the table. The distribu-
tions in Figure 7 reflect this adjustment. If larger intervals were to be specified
(e.g., 50 or 100 IUs) one might, based on the shape of the distribution, select to do
a similar chi-square goodness-of-fit comparison with a normal distribution.
We then proceeded to identify clusters by finding where the metaphor curve
deviated significantly from the Poisson curve. Working with 5 and 20 intona-
tion unit intervals, Poisson probabilities were calculated for the moving sums
of every consecutive interval (rather than the independent intervals needed
for the chi-square test), using a Poisson worksheet function in Microsoft Excel
spreadsheet software. We used probabilities of less than 0.05 as indicating a
cluster, as compared with the less conservative significance level of 0.10 used
by Corts and Pollio (1999). This statistical procedure returned clusters of two
sizes from the two interval lengths.
The Poisson procedure confirmed all the clusters marked on Figure 5. It
resolved the questions around E, F and G in Figure 6 by identifying only G as
a cluster. The Poisson procedure also confirmed the cluster around Jo’s reading
of her poem, identified in the cumulative frequency graph of the full discourse
event (Figure 4). Extract 1 (box B on Figure 5) was identified as a longer cluster,
and, inside it, the talk from lines 105 to 117, covering the first change of speaker,
was identified as a shorter cluster. Hence, combining the 5 IU and 20 IU interval
statistical analyses revealed clusters within longer clusters.
The first conversation was found to have 23 longer and 31 shorter metaphor
clusters. All but 6 of the shorter clusters were also part of a longer cluster. Two
of the longer clusters were combined with closely neighbouring clusters after
examination of the talk showed continuity of speaker and topic between them,
leaving 21 longer clusters. The second conversation, which was twice as long,
had 22 longer clusters and 68 shorter clusters. The mean length of the longer
clusters in the first conversation was 39 IUs, roughly equivalent to one and a
quarter minutes, with a minimum of 19 and a maximum of 122 IUs. The mean
length of clusters increased to 52 IUs in the second conversation, probably
because the participants were more relaxed with each other and because many
of the topics talked about were extended versions of those talked about in the
earlier meeting. Clusters included about 42 per cent of all metaphors while
accounting for 30 per cent of the talk. The rate of metaphor use increased on
average by 50 per cent inside clusters, as compared with non-cluster talk.
The statistical procedure offers researchers a reliable, if somewhat compli-
cated, method for identifying clusters of metaphors. The detailed work in using
the procedure makes it less amenable to exploratory work and researchers need
to decide whether the technical complexity of the procedure warrants its use.
130 Metaphor clusters in discourse

6.3 Advantages and limitations of using visual display to identify

metaphor clusters
The visual displays created using the VisDis software proved to be as accurate as
the statistical method in identifying metaphor clusters, and for many research
purposes will be ‘reliable enough’. The extra information about Vehicle domains
available in the visual display showed cluster boundaries more clearly than in
a cumulative frequency graph.
The facility for working with sub-categories of metaphor and for visualising a
range of discourse aspects on the same diagram makes the visual display useful
for exploring multiple co-occurrences and patterns.
The use of visualisation in the analysis of discourse is a fairly new venture.
Learning to ‘see’ and notice patterns in visualised data, and knowing how to inter-
pret them, require different discourse analytic skills from established, text-based
ones. A methodological research agenda will include investigation of the nature
of expertise in creating and interpreting visualisations of discourse data.

7 Metaphor clusters in conciliation talk

In this final section, we report some of our findings about the nature of meta-
phor clusters in the conciliation talk, comparing them with those reported in
the studies of lectures and sermons. We suggest how visual displays can be used
in investigating, as well as in identifying, metaphor clusters.

7.1 Quantitative comparisons with lectures and sermons

Table 2 summarises the occurrence of metaphor clusters in the two conciliation
conversations and in the studies of college lectures (Corts & Pollio, 1999), and
Baptist sermons (Corts & Meyers, 2002). With some of the more weak types
of metaphors omitted (Vehicle terms that are prepositions, or from thing and
part as lexemes), the conciliation conversations have a rate of metaphor use
similar to that of the sermons and higher than college lectures (first column of
Table 2). When the number of clusters produced in the three types of discourse
event are normed (second column of Table 2), the rate at which clusters are
produced in conciliation talk is again similar to that of the sermons, as is the
proportion of the total discourse event that occurs inside clusters, at around 30
per cent (third column of Table 2). However, the proportion of the total number
of metaphors used inside clusters (fourth column of Table 2), at around 41 per
cent, is more similar to lectures than to the much higher proportion in the
sermons. In other words, clusters are occurring at about the same rate across
the talk, but in the sermons, clusters are more densely packed with metaphors
than in the conciliation talk or the lectures.
 L. J. Cameron & J. H. Stelma 131

Metaphors Rate of % of total % of total

(figures) per production discourse metaphors
10 words / of clusters inside clusters (figures)
sentence (clusters per inside
Discourse type 100 minutes) clusters

Conciliation
conversations
1 0.7 a 40 30 42
2 0.7 a 20 29 41
Lectures b
1 0.3 6 11 50
2 0.3 9 6 33
3 0.3 8 6 33
Sermons c
1 0.6 25 29 74
2 0.3 40 20 71
3 0.8 30 43 87

a These figures exclude metaphors whose Vehicle terms are prepositions, thing or part as lexeme.
Including these increases the figure to 1.0
b from Corts and Pollio (1999)
c from Corts and Meyers (2002)

Table 2: Comparison of conciliation conversations, college lectures and sermons

8 The nature of metaphor clusters in the dynamics of

conciliation talk
8.1 Levels of clustering
The three methods used for identifying metaphor clusters produce a consistent
finding that metaphor clusters occur on several different levels or time scales
in the dynamics of discourse. Further, they all found certain levels emerging
as particularly relevant. These are:
• the overall discourse event (here, between 50 and 120 minutes);
• the 250 IU level (about 5 minutes of talk);
• the 20 IU level (about 25 seconds of talk);
• the 5 IU level (about 6 seconds of talk).
At the level of the overall event, analysts can see stages in talk, and the distribu-
tion of topics. At the 250 IU level, metaphor clusters become apparent in the
visual displays and frequency graphs. Metaphor clusters are empirically found
to occur at either the 20 IU level or the 5 IU level, with shorter clusters usually
found within larger clusters.
132 Metaphor clusters in discourse

9 The discourse work of metaphor clusters

The overall discourse purpose of conciliation is for the participants to reach
across the alterity between them in order to better understand each other’s
perspective. Clusters in the conversations were analysed for discourse purpose,
point of view presented, interaction between speakers, topic, and the connec-
tions between the metaphors in the cluster. Clusters did not, on the whole,
occur around changes in speaker or topic; only 3 of the 21 clusters included a
change of discourse topic (shown in the display by a horizontal line and change
of shading across a cluster), and only 3 clusters could be said to include jointly
constructed talk, rather than talk by one of the speakers with agreement or
minimal support utterances from the other.
It seems that participants used metaphors in clusters mostly to present their
own point of view to the Other. Twelve clusters involved the perspective of just
one participant, with nine including more than one perspective. There were 22
instances in clusters where a speaker presented his or her own point of view, 6
where they offered the Other’s point of view, and just 2 where a joint perspective
was presented. The presentation of one’s own point of view can be likened to the
expert explanations in lectures, classrooms and sermons, although it was often
more tentative. Different kinds of interactional work occur when a speaker
presents, again often tentatively, the Other’s point of view, or when together
they construct a shared perspective.
These three different dynamics of perspective (own, Other’s, joint) appear
differently on the visual display. In the alterity strand (see Figure 2), presenting
one’s own perspective will show as movement in the area on one side of the
centre column; presenting the Other’s perspective will show as larger zig-zags,
as the line moves backwards and forwards across the centre; the joint perspec-
tive will show as movement into the centre column of the strand. Finding
clusters with different perspective dynamics can be done visually by seeking
out these alterity patterns.
In contrast with the lectures and sermons, clusters in conciliation talk very
seldom arose from extensions of a single root or conceptual metaphor used
to explain an idea. For example, in the first conversation, all clusters bar one
included metaphors from more than one Vehicle domain. An extended root
metaphor appears on a visual display such as Figure 3b as a cluster of diamonds
in a single Vehicle domain column. Where clusters include more than one
extended metaphor, there are diamonds in more than one of the columns.
Two types of talk in clusters were unique to the interaction around concilia-
tion: appropriation of the Other’s metaphors and the exploration of alternative
scenarios. We illustrate these below.
 L. J. Cameron & J. H. Stelma 133

Within metaphor clusters at several key points in their conversations, each

of the two participants began to use a metaphor that had, until that point, been
the discourse ‘property’ of the Other. This could be seen as a significant step
towards acknowledging the Other’s perspective through using their language.
In Extract 2 below, the bomber makes first use of the metaphor of healing
to refer to himself. Before this, healing was Jo’s metaphor, used to talk about
recovering from grief. The appropriation of the metaphor is accompanied by
hesitation, pausing and permission seeking.

Extract 2: Appropriation of healing as metaphor

P: how do you put it, er,

…(2.0) maybe that’s part of healing too,
.. my healing
J: your healing.
.. [yeah]
P: [yeah]
…(1.0) you know,
er,
…(2.0) it’s –
something I have to go through
J: … hmh
P: .. if I’m going to sort of
er,
… (1.0) really retain my humanity

The topic of the healing metaphor shifts with this new use: Pat’s healing is to
do with accepting the consequences of his decision to plant the bomb, rather
than Jo’s recovering from grief. The appropriation of the metaphor makes use
of what Goatly (1997: 258) labels ‘multivalency’. The re-use of the Vehicle term
with part of, my, your as it is passed from one speaker to the other, contributes
to the formation of a cluster.
In the dynamics of the discourse, and using the vocabulary of complex sys-
tems theory, the metaphor cluster can be seen as the system moving to the ‘edge
of chaos’, where maximum flexibility of metaphor use reflects a shift in alterity
in the conciliation process. Alternatively, we can see the metaphor cluster as
marking a ‘critical’ moment in the discourse (Candlin, 1987).
A further discourse move that occurs within clusters in the conciliation talk is
when a speaker moves from what happened to what might have happened and
explores alternative actions that they could have taken. For Jo, the alternative
to seeking conciliation would have been to give into anger and revenge. Pat
several times contrasts acting from hatred with acting from political conviction.
134 Metaphor clusters in discourse

In Extract 3, Pat considers what motivated the IRA members that he knew, and
himself, and how this was in most cases something other than ‘hatred’.

Extract 3: Exploring the alternative: Being driven by hatred

2663 Pat I know very very few people,

2664 ...(1.0) that hatred was a big part of it.
2665 .. and those you could say to it,
2666 ... say that it was,
2667 ... wouldn’t have lasted.
2668 ... it’s –
2669 Jo [hmh]
2670 Pat [it’s] not enough to sustain,
2671 ... during a struggle like this.
2672 Jo hmh
2673 Pat .. I don’t think so.
2674 ... you –
2675 you couldn’t keep up with it,
2676 if it was just driven by that sort of –
2677 ...(1.0) hatred that gnaws away at you.
2678 Jo ... [hmh]
2679 Pat [you would] soon be a casualty of it.

The dense mixture of metaphors in IUs 2671–7 helps Pat to explore the alterna-
tive scenario in more depth and later to reject it. There were several instances of
this type of discourse activity from both speakers across the conversations, and
they seemed to be a significant and successful way of explaining to the Other
why each had made the choices they had made over the years, and how these
choices affected their lives. By enabling participants to verbalise what were often
very negative ‘might have been’ scenarios, they at the same time revealed to the
Other the strength of their feelings.

10 Conclusion
The unique features of metaphor clusters in conciliation talk, as well as features
shared with other discourse types, help to generalise from the various cases
studied so far. It appears that metaphor clusters occur when some intensive
interactional work linked to the overall purpose of the discourse is being
carried out. In lectures, lessons and sermons, the discourse work was mainly
explanations of difficult or unfamiliar topics, carried out through extended use
of one or two root metaphors. In conciliation talk, the central discourse work
of reducing alterity is pushed forwards through metaphor clusters, in which
 L. J. Cameron & J. H. Stelma 135

metaphors, as ‘ways of seeing one thing in terms of another’ (Burke, 1945), are
also tried out and gradually appropriated across speakers. Other clusters occur
as speakers explore alternative ways of feeling and acting, helping participants
contemplate and reject often very negative alternatives.
In extending what we know about the role of metaphor clusters and how to
identify them, this article offers researchers a heuristic tool to use in exploring
discourse. Faced with discourse data, a researcher can seek out clusters of
metaphors and investigate the interaction around and inside them, with the
expectations that the metaphor clusters indicate points where intensive and
important discourse work is carried out. Where large amounts of data are
involved, identifying and investigating clusters at the various levels of scale gives
a way of ‘cutting into’ data which can help understand the overall dynamics
of the discourse while at the same time identifying particular episodes worth
investigating in more detail.

Notes
1 The project, ‘Using visual display to investigate the dynamics of metaphor in concili-
ation talk’, was supported by the UK’s Arts and Humanities Research Board under its
Innovation Award scheme. We acknowledge that support, and also thank the partici-
pants in the talk for giving permission to use the data.
2 We were interested to note that their paper includes a visual display of root metaphors
by sentence that shares some features of our visual display, although we only became
aware of this after designing our display specifications.
3 Gesture, gaze and other paralinguistic features have not been analysed. The interaction
of gesture and metaphor has been investigated by McNeill (1992) and Cienki (1998), as
well as in the Corts and Pollio study (1999).
4 The ends of intonation units are marked with the following symbols:
, continuing intonation contour
. final intonation contour
– a truncated (incomplete) intonation unit
5 It would not be theoretically accurate to label these ‘conceptual metaphors’, since they
have emerged from a single discourse event, rather than across the language more
generally (in a Saussurean abstraction of langue from parole). Hence we use the term
‘systematic metaphor’ to highlight the systematicity of Topic and Vehicle domains across
the discourse.
6 The software was produced by ‘The Little Software Company’ as part of the project. An
evaluation copy can be obtained from the authors.
Any discrete feature of discourse that can be underlined in a transcript, not only meta-
phor, can be coded into the software and displayed.
136 Metaphor clusters in discourse

References
Burke, K. (1945) A Grammar of Motives. New York: Prentice Hall.
Cameron, L. (2003) Metaphor in Educational Discourse. London: Continuum.
Cameron, L. and Low, G. (2004) Figurative variation in episodes of educational talk and
text. European Journal of English Studies 8(3): 355–73.
Candlin, C. (1987) What happens when applied linguistics goes critical? In M. A. K.
Halliday, J. Gibbons and H. Nicholas (eds) Learning, Keeping and Using Language:
selected papers from the 8th World Congress of Applied Linguistics 461–86. Amsterdam:
John Benjamins.
Chafe, W. (1994) Discourse, Consciousness and Time. Chicago, IL: University of Chicago
Press.
Cienki, A. (1998). Metaphoric gestures and some of their relations to verbal metaphoric
expressions. In J.-P. Koenig (ed.) Discourse and Cognition: bridging the gap 189–204.
Stanford, CA: CSLI.
Corts, D. (1999) Spontaneous Production of Figurative Language and Gestures in College
Lectures: a comparison across disciplines. Knoxville: University of Tennessee.
Corts, D. and Meyers, K. (2002) Conceptual clusters in figurative language production.
Journal of Psycholinguistic Research 31(4): 391–408.
Corts, D. and Pollio, H. (1999) Spontaneous production of figurative language and gesture
in college lectures. Metaphor and Symbol 14(1): 8–100.
Goatly, A. (1997) The Language of Metaphors. London: Routledge.
Koller, V. (2003) Metaphor clusters, metaphor chains: analyzing the multifunctionality of
metaphor in text. Metaphorik 5: 115–34.
Lakoff, G. (1993) The contemporary theory of metaphor. In A. Ortony (ed.) (2nd edn)
Metaphor and Thought 202–251. New York: Cambridge University Press.
Low, G. (1997) A Celebration of Squid Sandwiches: figurative language and the manage-
ment of (non-core) academic text. Working paper. Language Teaching Centre, University
of York.
McCarthy, M. (1988) Some vocabulary patterns in conversation. In M. McCarthy (ed.)
Vocabulary and Language Teaching 181–200. London: Longman.
McNeill, D. (1992) Hand and Mind: what gestures reveal about thought. Chicago:
University of Chicago Press.
Pollio, H. and Barlow, J. M. (1975) A behavioural analysis of figurative language in psy-
chotherapy: one session in a single case study. Language and Speech 18: 236–54.
Stelma, J. and Cameron, L. (in preparation) Intonation units in research on spoken
interaction.
Thelen, E. and Smith, L. (1994) A Dynamic Systems Approach to the Development of
Cognition and Action. Cambridge, MA: MIT Press.
Wertsch, J. (1998) Mind as Action. New York: Oxford University Press.