D:/PhD/papers/DandD09incrementality/2011/2011-01 IncrementalitySplits.dvi


Dialogue and Discourse 2(1) (2011) 279–311 doi: 10.5087/dad.2011.111

On Incrementality in Dialogue:
Evidence from Compound Contributions∗

Christine Howes C.HOWES@QMUL .AC.UK

School of Electronic Engineering and Computer Science
Queen Mary University of London
Mile End Road, London E1 4NS, UK

Matthew Purver M .PURVER@QMUL .AC.UK

Queen Mary University of London

Patrick G. T. Healey PAT.HEALEY@EECS.QMUL .AC.UK

Queen Mary University of London

Gregory J. Mills GJMILLS@STANFORD.EDU

Department of Psychology
Stanford University
Stanford, CA 94305, USA

Eleni Gregoromichelaki ELENI.GREGOR@KCL .AC.UK

Philosophy Department
King’s College London
Strand, London WC2R 2LS, UK

Editor: Hannes Rieser and David Schlangen

Abstract

Spoken contributions in dialogue often continue or complete earlier contributions by either the
same or a different speaker. Thesecompound contributions(CCs) thus provide a natural context
for investigations of incremental processing in dialogue.

We present a corpus study which confirms that CCs are a key dialogue phenomenon: almost
20% of contributions fit our general definition of CCs, with nearly 3% being the cross-person case
most often studied. The results suggest that processing is word-by-word incremental, as splits
can occur within syntactic ‘constituents’; however, some systematic differences between same-
and cross-person cases indicate important dialogue-specific pragmatic effects. An experimental
study then investigates these effects by artificially introducing CCs into multi-party text dialogue.
Results suggest that CCs affect people’s expectations about who will speak next and whether other
participants have formed a coalition or ‘party’.

Together, these studies suggest that CCs require an incremental processing mechanism that
can provide a resource for constructing linguistic constituents that span multiple contributions and
multiple participants. They also suggest the need to model higher-level dialogue units that have
consequences for the organisation of turn-taking and for the development of a shared context.
Keywords: Compound contributions; corpus study; party formation; dialogue; incrementality

∗. This research was carried out under theDynamics of Conversational Dialogueproject, funded by the UK
ESRC (RES-062-23-0962) with EPSRC support through theDiET (Dialogue Experimentation Toolkit)project
(EP/D057426/1). We also thank Ruth Kempson and Arash Eshghi for many useful discussions.

c©2011 C. Howes, M.Purver, P. G. T. Healey, G. J. Mills and E. Gregoromichelaki Submitted 1/10; Accepted 4/11; Published online 5/11



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

1. Introduction

Compound contributions(CCs) – spoken dialogue contributions that continue or complete an earlier
contribution,1 see e.g. (1) – have been claimed to occur regularly in dialogue, especially according
to the Conversation Analysis (CA) literature, where specific types of compound contributions have
been studied under a variety of names, including completions and joint productions (see section 2).

(1) Daughter: Oh here dad, a good way to get those corners out

Dad: is to stick yer finger inside.

Daughter: well, thats one way. [from Lerner (1991)]

CCs are of interest to dialogue theorists as they provide evidence about how contributions can
cohere with each other at multiple levels – syntactic, semantic and pragmatic (though of course
they are not the only way). They also indicate the radical context-dependency of conversational
contributions, which can, in general, be highly elliptical without disrupting theflow of the dialogue.
CCs are a dramatic illustration of this: speakers must rely on the dynamics of theunfolding context
(linguistic and extra-linguistic) in order to guarantee successful processing and production.

As early as 1967, in his series of Lectures on Conversation, Sacks (1992) noted that the ex-
istence of CCs supports the (now largely accepted) thesis that language indialogue is processed
incrementally:

Such a fact as that persons go about finishing incomplete sentences of others with
syntactically coherent parts would seem to constitute direct evidence of their analysing
an utterance syntactically in its course. . . (Sacks, 1992, p651)

However, we argue here that the evidence from CCs goes further; they show that not just processing
(parsing), but also production (generation) must be incremental; and that because of the variation in
CCs, this must also be at a finer-grained level than is often assumed (see also Ferreira, 1996; Guhe,
2007).

Compound contributions that are split across speakers also present a canonical example of par-
ticipant coordination in dialogue (here we call thesecross-person CCsto distinguish them from the
same-personcases where the original speaker later continues his own contribution – see below).
The ability of one participant to continue another interlocutor’s contribution coherently, both at the
syntactic and semantic level, implies that speaker and hearer can be highly coordinated in terms of
processing and production. The initial speaker must be able to switch to the role of hearer, pro-
cessing and integrating the continuation of their contribution, whereas the initial hearer must be
monitoring the grammar and content of what they are being offered closely enough that they can
take over and continue in a way that respects the constraints set up by the first contribution. This
switch is particularly obvious in those cases where the initial hearers continuation is not the same
as that which the original speaker would have provided, as in (1, 2).

1. These terms will be defined in detail below.

280



COMPOUND CONTRIBUTIONS

(2) BMA: She got compensation

Just like that

Because what she had in her suitcase

PM: was Grade A.

[from comedy news quizHave I got news for you, s35 ep1]

There is evidence that such constraints are respected across speaker and hearer in compound
contributions (see e.g. Gregoromichelaki et al., 2009). In both Finnish (which has a rich inflectional
morphology), and Japanese (a verb-final language), cross-person CCs within a single clause con-
form to the strict syntactic constraints of the language, despite the change inspeaker (Helasvuo,
2004; Hayashi, 1999; Lerner and Takagi, 1999).

These observations have important theoretical implications. Firstly, the grammar and seman-
tics employed by the interlocutors must be able to license and interpret chunksmuch smaller than
the usual sentential or propositional units. Moreover, the possibility of role switches while syntac-
tic/semantic dependencies are pending suggests direct involvement of the grammar in the parsing
and production processes, or, at least, a very tight coupling between those processes and the gram-
mar and intermediate representations being used (see Gargett et al., 2009). Indeed, Poesio and
Rieser (2010) claim that “[c]ollaborative completions . . . are among the strongest evidence yet for
the argument that dialogue requirescoordinationeven at the sub-sentential level” (italics original).

From a psycholinguistic point of view, the phenomenon of CCs is compatible withmechanistic
approaches as exemplified by the Interactive Alignment model of Pickeringand Garrod (2004),
which claims that it should be as easy to complete someone else’s sentence as one’s own (p186).
According to this model, speaker and listener ought to be interchangeable at any point. This is
also the stance taken by the grammatical framework of Dynamic Syntax (DS: Kempson et al., 2001;
Cann et al., 2005). In DS, parsing and production are taken to employ the same mechanisms, leading
to a prediction that CCs ought to be strikingly natural (Purver et al., 2006).However, continuation
by another speaker is sometimes taken to involve guessing or preempting the other interlocutor’s
intended content.2 It has therefore been claimed that a full account of CCs requires a complete
model of pragmatics that can handle intention recognition and formation. Indeed, Poesio and Rieser
(2010) propose sentence completions as the testing ground of competing claims about coordination
i.e. whether it is best explained with an intentional model like Clark’s (1996) or with a model based
on simpler alignment models like Pickering and Garrod’s (2004). They conclude that a model which
includes modelling of intentions better captures the data, though see (Gregoromichelaki et al., 2011)
for an alternative argument.

For computational models of dialogue, compound contributions pose a challenge. While Poesio
and Rieser (2010) and Purver et al. (2006) provide general foundational models for various aspects
of CCs, there are many questions that remain if automatic processing of naturally occurring dia-
logues is ever to be completely realised. A computational dialogue system must be able to identify
CCs, match up their two (or more) parts (which may not necessarily be adjacent), integrate them
into some suitable syntactic and/or semantic representation, and determine the overall pragmatic
contribution to the dialogue context. CCs also have implications for the organisation of turn-taking
in such models (see e.g. Sacks et al., 1974), as regards what conditions(if any) allow or prevent
successful turn transfer.

2. Note that this says nothing about whether such a continuation is the same as the initial speaker’s intended continua-
tion. For cases where this cannot be the case see (Gregoromichelaki et al., 2011), as well as (1, 2) above.

281



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

From an organisational point of view, it has been claimed that turn-taking operates not on indi-
vidual conversational participants, but on ‘parties’ (Schegloff, 1995). For example, a couple talking
to a third person may organise their turns as if they are one ‘party’, ratherthan two separate individ-
uals. Lerner (1991) speculates that cross-person compound contributions can clarify the formation
of such parties, as they reveal a relationship between syntactic mechanismsand social organisa-
tion. He claims that this provides evidence of one way in which syntax can be used to organise
participants into “groups”.

Of course, our earlier definition of compound contributions begs several questions; most impor-
tantly, what do we mean by a ‘dialogue contribution’? Our use of this term, andthe related notion
of a turn, can be best explained by reference to a short extract of dialogue taken from the British
National Corpus (3).

(3)
1. A: I were gonna say, they wash [[better than]]
2. J: [[But I’ve had]]
3. A: velvet.
4. J: I’ve had to take them up.
5. Cos they were, they were gonna be miles too long.
6. And I’ve not even took them out the thing.
7. They said he’d swap them if they didn’t fit.
8. A: [[Ah they do!]]
9. J: [[And he]]<pause><unclear>.
10. A: Where d’ya get them from Joyce?
11. J: I got them from that er
12. B: Top Marks.
13. J: that shop. [BNC KB2 4134-4146]

In our usage, each of the transcribed lines (1-13) is acontribution. Our use ofcontribution is
intended to correspond to Clark’s (1996) “acontribution to discourse – [is] a signal successfully
understood” (p227).3 With transcribed corpus text, of course, it is not always possible to deter-
mine whether contributions have been successfully understood, as we have no access to non-verbal
signals (such as nodding). We therefore take contributions to be stretches of talk bounded by a
change in speaker, a significant pause, or the end of the sentence, and assume that in most cases
the transcribers’ decision to split the text into separate lines indicate some (e.g. prosodic) cues to
suggest that the line has been successfully understood, i.e. treated as acontribution. Thus, whilst
contributions can be single words (as in line 3) or backchannels (e.g. ‘mm’), or complete syntactic
sentences (e.g. line 4), they can also be partial sentences (e.g. the incomplete sentences at lines
1, 2 and 11 and the fragments at lines 3, 12 and 13). Note however, that single words in longer
contributions (e.g. ‘they’ at the start of line 7) do not count as contributions in their own right.

Compound contributionscan now be defined assingle syntactic or semantic (propositional)
units built across multiple contributions, which could be provided by one speaker or several. The
exchange in lines 11-13 provides two examples. J’s contribution ‘I got them from that er’ starts a
sentence, which B’s contribution ‘Top Marks’ (the name of a shop) completes. This counts as a

3. Note that Clark uses contribution to refer to both “the joint act of . . . completing the signal and its joint construal”
and for the interlocutors “participatory act, hispart of that joint act, as when we speak of Roger’s contribution to the
discourse.” We use contribution in this second sense only.

282



COMPOUND CONTRIBUTIONS

compound contribution under our definition. J then also completes her own contribution (with ‘that
shop’) at line 13, and this also counts as a (same-person) compound contribution, as it is spread
across multiple contributions (in this case, with intervening material). Note that even though the
short extract in (3) also exhibits many other conversationaltying techniques(Sacks, 1992), such as a
question and answer (lines 10-11), and the use of pronouns linked to referents previously introduced
in the dialogue, our focus here is not on all pragmatic dependencies between turns.

It should be noted, however, that this definition depends on the protocolused by the corpus
transcribers; and with the BNC, this can lead to possibly undesirable segmentation of stretches of
talk into multiple “contributions”. The insistence on linear ordering means that cases of interruption
of one speaker by another will always result in an apparent speakerchange, even if the interruption
consists only of non-verbal noises (e.g. coughing) or is entirely overlapping – see e.g. lines 1-3
(overlapping material is shown in the examples with square brackets aligned tothe material with
which it overlaps). J’s interruption in line 2 overlaps with A’s speech, butforces A’s sentence to be
transcribed as two lines (1 and 3). These count as separate contributions under our definition, giving
a compound contribution: A begins her contribution ‘I were gonna say, they wash better than’, which
she completes in line 3 with ‘velvet’. In many cases this may be the correct analysis – in Clark’s
usage, overlapcansignal understanding (I might not need you to syntactically or semantically finish
your sentence to accept it as a valid contribution to the discourse). In this case, though, it may be that
lines 1 and 3 were intended (and processed) as one single contribution – toavoid possibly misleading
conclusions we therefore report CC figures both including and excluding such cases (see section 4).
Note, however, that these concerns only apply to same-person CCs andnot to cross-person CCs.

Whe also define a notion ofturn here as all talk to the next change of speaker; the contributions
by J in lines 4-7 would therefore be classified as a singleturn. We will use this notion below
to distinguish CCs which span multiple turns from those spanning multiple contributions within a
single turn. Even a backchannel or overlapping material, such as line 2 (which completely overlaps
with the end of line 1) counts as a change of speaker (and thus separate turns) here.

Analysis of CCs, when they can or cannot occur, and what effects they have on the coordination
of agents in dialogue, is therefore an area of interest not only for conversation analysts wishing to
characterise systematic interactions in dialogue, but also for linguists trying toformulate grammars
of dialogue, psychologists and sociolinguists interested in alignment mechanisms and social inter-
action, and those interested in building automatic dialogue processing systems.In this paper we
present and examine empirical corpus data and an experimental manipulationof CCs, in order to
shed light on some of the questions and controversies around this phenomenon.

2. Related Work

Most previous work on what we call CCs has examined specific sub-cases, generally of the cross-
person type, and have referred to these variously ascollaborative turn sequences(Lerner, 1996,
2004),collaborative completions(Clark, 1996; Poesio and Rieser, 2010),co-constructions(Sacks,
1992),joint productions(Helasvuo, 2004),co-participant completions(Hayashi 1999, Lerner and
Takagi 1999),collaborative productions(Szczepek, 2000a) andanticipatory completions(Fox,
2007) amongst others (with some differences of emphasis in the differentterms). Here we discuss
some of this work.

283



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

2.1 Conversation Analysis

Anticipatory Completions Lerner (1991) identifies various structures typical of CCs which con-
tain characteristic split points. One group of these are ‘compound’turn-constructional units(TCUs),
which are structures that include an initial constituent that hearers can identify as introducing some
later final component. Examples include theIF X-THEN Y, WHEN X-THEN Y and INSTEAD OF

X-Y constructions (4).

(4) A: Before that then if they were ill

G: They get nothing. [BNC H5H 110-111]

Other cues for potentialanticipatory completionsinclude quotation markers (e.g.SHE SAID), par-
enthetical inserts and lists, as well as non-syntactic cues such as contrast stress or prefaced disagree-
ments. Another important category that he identifies isterminal itemcompletions, which involve
completing the final one or two lexical items of an interlocutor’s utterance at projectable locations
of the current speaker’s turn ending (possibly involving overlap).

Opportunistic Cases Although Lerner focuses on these projectable turn completions, he also
mentions that CCs can occur at other points such as “intra-turn silence”, laugh tokens and hes-
itations, for example in cases of a stalled word search. All these cases he terms opportunistic
completions(5).

(5) A: Well I do know last week thet=uh Al was certainly very〈pause 0.5〉

B: pissed off [Lerner (1996), p260]

As he makes no claims regarding the frequency of such devices for CCs,it would be interesting to
know how common these are, especially as studies on CCs in Japanese (Hayashi, 1999) show that
although CCs do occur, compound TCUs do not play as prominent a role asin English. It should be
noted, however, that Lerner’s definitions are not intended to be mutually exclusive.

Expansions vs. Completions Other classifications of CCs often distinguish betweenexpansions
and completions(Ono and Thompson, 1993). Expansions are continuations which add, e.g., an
adjunct, to an already complete syntactic element (6, 7).

(6) T: It’ll be an E sharp.

G: Which will of course just be played as an F. [BNC G3V 262-263]

(7) M: yep dr goes everyones happy

N: except the dr [DiET SU1 4213-4214]

Completions involve the addition of syntactic material which is required to make the whole com-
pound contribution (syntactically) complete (5, 8).

(8) A: . . . and then we looked along one deck, we were high up, and down belowthere were
rows of, rows of lifeboats in case you see

B: There was an accident.

A: of an accident [BNC HDK 63-65]

284



COMPOUND CONTRIBUTIONS

Importantly, though we consider both expansions and completions to be CCs according to our
terminology, we distinguish between the two types by considering the completeness or otherwise
of the first part of the CC. Thus while there might be arguments for restricting the definition of a
CC to only the completion type, we are also interested in comparing the relative distributions of the
different sub-types.

In terms of frequency, the only estimate we are aware of in the CA literature is Szczepek (2000a),
who found approximately 200 cross-person CCs in 40 hours of English conversation (there is no
mention of the number of sentences or turns this equates to), of which 75% are completions.

As briefly outlined above, CA analyses of CCs tend to focus on their sequential implications in
particular cases. These analyses provide clear examples of cross person co-ordination, however, it
is unclear how representative they are (with the exception of Szczepek (2000a), who offers limited
figures). Additionally, as the emphasis in the CA literature on CCs is in identifyingtheir organi-
sational consequences for the unfolding dialogue (which can range from indicating understanding
to highlighting differences of opinion (Szczepek, 2000b)), they leave open the question of where a
speaker switch may occur.

2.2 Linguistic Models

Purver et al. (2006) present a grammatical model for compound contributions, using an inherently
incremental grammar formalism, Dynamic Syntax (Kempson et al., 2001; Cann et al., 2005). This
model shows how syntactic and semantic processing can be accounted forno matter where the
split point occurs; however, as their interest is in grammatical processing, they give no account of
any higher-level inferences which may be required. Poesio and Rieser(2010) present a general
model forcollaborative completions(a subclass of cross-person CCs) based in the PTT framework,
using an incremental LTAG-based grammar and an information-state-basedapproach to context
modelling. While many parts of their model are compatible with a simple alignment-based com-
munication model like Pickering and Garrod’s (2004), they see intention recognition as crucial to
dialogue management. They conclude that an intention-based model, like Clark’s (1996), is more
suitable. Their primary concern is to show how such a model can account for the hearer’s ability
to infer a suitable continuation, but their use of an incremental interpretation method also allows
an explanation of the low-level utterance processing required. Nevertheless, the use of an essen-
tially head-driven grammar formalism suggests that some syntactic split points ought to be more
problematic than others.

2.3 Corpus Studies

Skuplik (1999) collected data from German two-party task-oriented dialogue, and annotated for
cross-person compound contribution phenomena. She found thatexpansions(cases where the part
before the split point can be considered already complete, as describedabove) were more common
thancompletions(where the first part is syntactically or semantically incomplete as it stands), with
72 expansions (57%) and 54 completion CCs (43%) in her corpus. This contrasts with the data
reported by Szczepek (2000a), detailed above. There are severalpossible reasons for this contrast;
for example, there may simply be a difference in the distributions of CCs in different languages, or
between experimentally controlled task-oriented dialogue (which Skuplik (1999) focused on) and
casual conversational dialogue. Additionally, there may be issues with the classification schemes
used. For example, Szczepek (2000a) did not include what she callsappendor questionsin her data,

285



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

which could also be argued to be expansion CCs. The corpus study hereshould shed some light on
some of these possible sources of disagreement.

Rühlemann (2007) uses corpus analysis on the BNC to examine a subset ofexpansionCCs,
sentence relativesof one’s own or another’s turn (6, 9).

(9) A: profit for the group is a hundred and ninety thousand pounds.

B: Which is superb. [BNC FUK 2460-2461]

He found thatsentence relativesare slightly more likely to be same-person than cross-person, with
a total of 104 (55%) of 190 being same-person cases. This contrasts withTao and McCarthy (2001)
who found 96% of their corpus sample were same-person; however, thisdiscrepancy can be at-
tributed to the fact that they were measuring different things: Tao and McCarthy (2001) included all
non-restrictive (‘which’) relative clauses in their analysis, thus excluding restrictive readings, and
including cases which were intra-sentential and thus would not count as CCs in our terminology (see
section 3.1). In fact, R̈uhlemann (2007) also excluded intra-turn cases where the sentence relative
was annotated as a separate sentence but there was no intervening material; our definition would
include these.

In addition, de Ruiter and van Dienst (in preparation) are also in the process of studying cross-
personcompletionsand their effect on the progressivity of dialogue turns; however no results are
available to us at this point in time. Notably, the definition used by de Ruiter and van Dienst (pc)
only includes those completions where the additional material combines with the incomplete first
part of the CC such that neither part could be considered complete withoutthe other. In our view,
this excludes a number of interesting cases; not only expansion type CCs,but also those in which
the continuation does not finish in a complete way (including, for example, CCswhich spread over
more than two parts).

2.4 Dialogue Models

Skantze and Schlangen (2009) and Buß et al. (2010) present incremental dialogue systems (for
limited domains) which can deal with some kinds of same-person compound contribution, allowing
the system or user to provide mid-sentence backchannels, and/or resumewith sentence completion
if interrupted. Some related empirical work regarding the issue of turn-switch addressed here is
also presented by Schlangen (2006) but the emphasis there centers mostlyon prosodic rather than
grammar/theory-based factors.

For cross-person CCs, the only system we are aware of is that presented in DeVault et al. (2009)
in which the system is able to generate a completion to a user’s input based on the semantic rep-
resentation it has built up so far. Due to the limited domain of possible semantic interpretations,
the system is able to produce terminal item completions, once the possible interpretations have been
sufficiently narrowed down. It does not, therefore, produce the range of CCs seen in naturally occur-
ring human dialogue (including expansions as discussed above); we hope that empirical data such
as that presented here can be used in constructing such systems and evaluating whether they achieve
DeVault et al.’s stated aim of enabling virtual agents to display natural conversational behaviour.

286



COMPOUND CONTRIBUTIONS

3. General Methods

3.1 Terminology

In this paper, as our interest is general, we use the termcompound contributions (CCs) to cover all
instances where more than one dialogue contribution combine to form a (intuitively propositional)
unit – whether the contributions are by the same or different speakers. We therefore use the term
split point to refer to the point at which the compound contribution is split (rather than e.g. transi-
tion pointwhich is associated with a speaker change). Cases where the speaker does change across
the split point are calledcross-person CCs; otherwise we call themsame-person CCs.

As not all cases will lead to complete propositions, and not all will be split over exactly two
contributions, we also avoid terms likefirst-half, second-halfandcompletion: instead the contribu-
tions on either side of a split point will be referred to as theantecedent and thecontinuation. In
cases where an compound contribution has more than one split point, some portions may therefore
act as the continuation for one split point, and the antecedent for the next.We can then talk about
completeness of each portion independently, with the traditional completion/expansion distinction
corresponding to completeness (or otherwise) of the antecedent. See thesub-section on annotation
scheme in section 4.1 for details of how completeness is assessed.

3.2 Questions

Questions about frequencies and distributions are addressed in the corpus study (section 4); these
lead to others about the effects on the ongoing dialogue, which are examined in the experimental
manipulation (section 5).

General Our first interest is in the general statistics regarding CCs: how often do they occur?
When they do, do they usually fall into the specific categories (with specific preferred split points)
examined by e.g. Lerner (1991), or can the split point be anywhere? What effects do CCs then have
on the ongoing dialogue? Do same- and cross- person CCs have different effects? Specifically, do
CCs have a bearing on ‘party’-formation in Schegloff’s (1995) sense, as Lerner (1991) claims?

Same- vs cross-person We are also interested in the balance between same- and cross-person
CCs. Some grammatical formalisms (Purver et al., 2006) and psycholinguistic models (Pickering
and Garrod, 2004) predict that CCs should be equally natural in both same- and cross- person
conditions – is this the case? What are the similarities and differences betweensame- and cross-
person cases?

Completeness For a grammatical treatment of CCs, as well as for implementing parsing/produc-
tion mechanisms for their processing, we need to know about the likely completeness of antecedent
and continuation (for example, if they are always complete in their own right, astandard head-driven
grammar may be suitable; if not, something more fundamentally incremental may be required). In
addition, CA analyses of dialogue phenomena predict that compound contributions should prefer-
ably occur at turn-transfer points that are foreseeable by the participants. Complete syntactic units
serve this purpose from this point of view and lack of such completeness will seem to weaken this
general claim.

We therefore ask how often antecedents and continuations are themselvescomplete. For an-
tecedents, we are more interested in whether theyend in a way that seems complete as they may
have started irregularly due to overlap or another CC (end-complete); for continuations, whether

287



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

theystart in such a way – they may not get finished for some other reason, but we want to know if
they would be complete if they do get finished (start-complete).4 These notions are by no means
entirely clear cut (as pointed out by an anonymous reviewer there is much debate on whether e.g.
adverbial adjuncts and semantic roles are necessary in a sentence) andSchegloff (1996) concedes
that his definitions are both arguable and not fully specified, although conversational participants do
orient themselves to points ofpossiblecompletion. In practice, however, in most cases there was
a high level of agreement between annotators on what constitutes syntacticor semantic complete-
ness.5 We also look at the syntactic and lexical categories which occur either side of the split point.
We are interested to know whether there are different effects on the unfolding dialogue from CCs
with complete and incomplete antecedent contributions, and whether the positionof the split point
has an effect.

Repair and Overlap Finally, we look at how often the continuation of an CC involves explicit
repair (repetition, reformulation, modification or replacement) of antecedent material. Any gram-
mar of dialogue or computational system will need to be able to identify where thistakes place, and
we therefore also look at how such repair depends on antecedent completeness and the type of split
point.

As our focus is on CCs, note that our use ofrepair refers only to those cases where the ‘end’
of the antecedent (immediately preceding the split point) is explicitly repeated orreframed at the
start of the continuation. An example can be seen in (12), where the last word of the antecedent
is repeated in the continuation. Repairs at other points in the s-unit or turn are not taken into
consideration.6

4. Study 1: Corpus Study

4.1 Materials and Procedure

For this exercise we used the portion of the BNC (Burnard, 2000) annotated by Ferńandez and
Ginzburg (2002), chosen to maintain a balance between what the BNC defines as context-governed
dialogue (tutorials, meetings, doctor’s appointments etc.) and demographic dialogue (casual un-
planned conversations). This portion comprises 11,469s-units– roughly equivalent to sentences7 –
taken from 200-turn sections of 53 separate dialogues.

The BNC transcripts are already annotated for overlapping speech, for non-verbal noises (laugh-
ter, coughing etc.) and for significant pauses. Punctuation is included, based on the original audio
and the transcribers’ judgements; as the audio is not available, we allowed annotators to use punc-

4. The notion of end-completeness that we are trying to capture is the CA notion of endingsas outlined in Sche-
gloff (1996); “for any TCU we can ask . . . does it end with an ending, i.e., does it come to a recognizable possible
completion – syntactic, prosodic and action/pragmatic.” Likewise hisbeginningsfor our start-completeness; “Turn
constructional units – and turns – can start with a “beginning” or with something which is hearablynot a beginning.”

5. See the sub-section on annotation scheme in section 4.1 for operational details, and table 2 for kappa agreement
scores between annotators.

6. Consequently, our use ofrepair should be understood not as capturing all instances of repair but only as indexing
the frequency with which these specific aspects of the contribution are repaired.

7. The BNC is annotated intos-units, defined as “sentence-like divisions of a text”, andutterances, defined as “stretches
of speech usually preceded and followed by silence or by a change of speaker”. Utterances may consist of many
s-units; s-units may not extend across utterance boundaries. While s-units are therefore often equivalent to complete
syntactic sentences, or complete functional units such as bare fragments or one-word utterances, they need not be:
they may be divided by interrupting or overlapping material from anotherspeaker.

288



COMPOUND CONTRIBUTIONS

tuation where it aided interpretation. The BNC transcription protocol divides the transcript into
sentence-like units (“s-units” ) as well as speaker turns (“utterances” – see footnote 7), where ut-
terances may contain several s-units from the same speaker. We annotated at the level of individual
s-units, to allow self-continuations within a turn to be examined; we are therefore taking the BNC’s
s-unit to correspond to our notion of dialoguecontribution, and the BNC’sutteranceas our notion
of turn.

The BNC forces speaker turns to be presented in linear order, which is vital if we are to ac-
curately assess whether turns are continuations of one another; however, this has a side-effect of
forcing long turns to appear as several shorter turns when interruptedby intervening backchannels.
We will discuss this further below.

Tag Value Explanation
end-complete y/n For all s-units: does this s-unit end in such a way as to yield a complete

proposition or speech act?
continues s-unit ID For all s-units: does this s-unit continue the proposition or speech act of

a previous s-unit? If so, which one?
repairs number of words For continuations: does the start of this continuation explicitly repair

words from the end of the antecedent? If so, how many?
start-complete y/n For continuations: does this continuation start in such a way as to be

able to stand alone as a complete proposition or speech act?

Table 1: Annotation Tags

Annotation Scheme The initial stage of manual annotation involved four tags:end-complete,
continues, repairs and start-complete – these are explained in Table 1 above. S-
units which somehowrequire continuation (whether they receive it or not) are therefore those
markedend-complete=n; s-units which act as continuations are those marked with non-empty
continues tags; and their antecedents are the values of thosecontinues tags. Further spe-
cific information about the syntactic or lexical nature of antecedent or continuation could then be
extracted (semi-) automatically, using the BNC transcript and part-of-speech annotations.

289



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

(3)

e
n
d
-
c
o
m
p
l
e
t
e

c
o
n
t
i
n
u
e
s

r
e
p
a
i
r
s

s
t
a
r
t
-
c
o
m
p
l
e
t
e

1. A: I were gonna say, they wash [[better than]] n
2. J: [[But I’ve had]] n
3. A: velvet. y 1 n
4. J: I’ve had to take them up. y 2 3 y
5. Cos they were, they were gonna be miles too long. y 4 n
6. And I’ve not even took them out the thing. y 5 n
7. They said he’d swap them if they didn’t fit. y
8. A: [[Ah they do!]] y
9. J: [[And he]]<pause><unclear>. y 7 n
10. A: Where d’ya get them from Joyce? y
11. J: I got them from that er n
12. B: Top Marks. y 11 n
13. J: that shop. y 11 1 n

Returning to the extract in (3), repeated here, we can see how these tagsare applied in practice.
Note that all s-units have anend-complete tag whilst only those that are judged to continue
some prior contribution have any other tags. The reason for judging end-completeness rather than
whether the s-unit constitutes a complete proposition or speech act in its own right, is due to both
the fragmentary nature of dialogue and the transcription practices of the BNC, which, as already
discussed, may break up a syntactic sentence into several s-units due to overlapping material etc.
Whether an s-unitendsin a potentially complete way is therefore independent of whether itstarts
in one. For thecontinues tag, the value is the line number which this s-unit is judged to continue
(i.e. the line number of the antecedent); lines 12 and 13, for example, are both judged to be a
continuation of line 11. Therepair tag takes as its value (if it has one) the number of words from
the end of the antecedent which are repeated, reformulated, modified or replaced at the start of the
continuation. Line 4 has arepair value of 3, because the continuation repeats the three words
from the end of line 2 (which is the antecedent) –‘I’ve had’.8 Finally, thestart-complete tag
(also only applied to continuations) indicates whether the contribution starts in away that it might
be the beginning of a complete sentence (even though it may not itself be complete). Continuations
starting withand/or/but/becauseetc. are always tagged asstart-complete=n, as can be seen
in lines 5, 6 and 9.

Inter-Annotator Agreement In some cases, it is not easy to identify whether a fragment is a
continuation or not, or what its antecedent is – see e.g. (10), where G’s second contribution could
be seen as continuing either his own prior utterance, or A’s intervening contribution:

(10) G: Well a chain locker is where all the spare chain used to like coil up

A: So it 〈unclear〉 came in and it went round

8. ‘I’ve’ is counted as two words as a contraction of‘I have’.

290



COMPOUND CONTRIBUTIONS

G: round the barrel about three times round the barrel then right down into the chain locker
but if you kept, let it ride what we used to call let it ride well〈unclear〉 well now it get so
big then you have to run it all off cos you had one lever, that’s what you had and the
steam valve could have all steamed. [BNC H5G 174:176]

Similar issues also arise in judgements of completeness, as it is not always obvious if a contribution
is syntactically or semanticallyend- and/orstart-complete. We therefore assessed inter-
annotator agreement between the three authors who acted as annotators.First, all three annotated
one dialogue independently, then compared results and discussed differences. They then annotated
3 further dialogues independently and agreement was measured; kappastatistics (Carletta, 1996)
are shown in Table 2 below.

BNC Dialogue Code
Tag KND KBG KB0

end-complete .86-.92 .80-1.0 .73-.90
continues (y/n) .81-.89 .76-.85 .77-.89
continues (ant) .82-.90 .74-.85 .76-.86

repairs 1.0-1.0 .55-.81 1.0-1.0
start-complete .59 .68 .62

Table 2: Inter-Annotatorκ statistic (min-max)

With the exception of therepairs tag for one annotator pair for one dialogue and thestart-
complete tags, all are above 0.7; the low figure in therepair category results from a few
disagreements in a dialogue with only a very small number ofrepairs instances. Thestart-
complete kappa figures, between the two annotators who completed this task, are around 0.6
suggesting that this measure may be less easy to determine. The remaining dialogues were then
divided evenly between the three annotators.

4.2 Results and Discussion

The 11,469 s-units annotated yielded 2,231 CCs, of which 1,902 were same-person and 329 cross-
person cases; 112 examples involved an explicit repair by the continuationof the antecedent. The
data come from the full range of dialogues; all dialogues had at least three same-person cases,
though 5 of the 53 dialogues had no cross-person CCs. The mean numberof same-person CCs
is 35.89 per dialogue (standard deviation 22.46). For cross-person CCs the mean was 6.21 per
dialogue (s.d. 5.69).

Within- and cross-turn cases Same-person CCs are much more common than cross-person; how-
ever, many of these same-person cases (around 44%) are self-continuations within a single speaker
turn (such as those between lines 4 and 5 in (3)). As explained in section 3.2, we consider same-
person cases to be interesting in their own right. From a processing/psycholinguistic point of view,
we would like to know whether such split points occur in the same places in cross-person CCs as
in same-person CCs. However, there are certainly arguments for considering CCs within a turn
as single contributions, and including them when comparing the frequency or nature of same- and
cross-person CCs may give an unfair comparison, as cross-personCCs can only occur at speaker
turn boundaries.

291



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

In addition, some apparently cross-turn cases (around 17%) may in factonly appear as such
due to the BNC transcription protocol, which forces speaker turns to be strictly linearly ordered.
A sentence from a single speaker which is interrupted by material from another speaker will be
transcribed as two separate turns – even if the intervening material is non-verbal (e.g. a cough)
and/or entirely overlaps with the original sentence rather than actually interrupting its flow (as seen
in (3) lines 1-3). In the tables and results below, we therefore present same-person CC figures both
including all cases, and excluding those cases which are either within-turnor separated only by
non-verbal or overlapping material. We label these figures asall andcross-turnrespectively.

person: Same- Cross-
all cross-turn (all)
N % N % N %

overlapping 0 0 0 0 18 5
adjacent 840 44 0 0 262 80
sep. by overlap 320 17 0 0 10 3
sep. by backchnl 460 24 456 63 17 5
sep. by 1 s-unit 239 13 229 32 16 5
sep. by 2 s-units 31 2 31 4 4 1
sep. by 3 s-units 5 0 3 0 1 0
sep. by 4 s-units 4 0 4 1 0 0
sep. by 5 s-units 1 0 1 0 0 0
sep. by 6 s-units 2 0 2 0 1 0
Total 1902 726 329

Table 3: Antecedent/continuation separation

General Observations Looking at cross-turn cases, even excluding those within-turn and over-
lapping cases discussed above, there are over twice as many same-person CCs (726) as cross-person
CCs (329). Many CCs have at least one s-unit intervening between the antecedent and continuation
(see Table 3). In same-person cases, once we have excluded the within-turn CCs described above,
this must in fact always be the case (see, for example, lines 11 and 13 in (3), where the contribution
at line 12 means that the antecedent (line 11) and continuation (line 13) are non-adjacent); the in-
tervening material is usually a backchannel (63% of remaining cases) or asingle other s-unit (32%,
often e.g. a clarification question), but two intervening s-units are possible(4%) with up to six being
seen. In cross-person cases, 88% are adjacent or separated onlyby overlapping material, but again
up to six intervening s-units were seen, with a single s-unit most common (10%,in half of which
the intervening s-unit was a backchannel).

Many compound contributions have more than two separate contributions. Insame-person
cases, a CC can be split over as many as thirteen individual s-units; although such extreme cases
occur generally within one-sided dialogues such as tutorials, many multi-split cases are also seen
in general conversation. Only 63% of cases consisted of only two s-units. Antecedents can also
receive more than one competing continuation (as in (3), where line 11 is continued in both lines 12
and 13), although this is rare: two continuations are seen in 2% of cases.

CA Categories We searched for examples which match CA categories (Lerner, 1991; Rühlemann,
2007) by looking for particular lexical items on either side of the split point. This search was per-

292



COMPOUND CONTRIBUTIONS

formed in two stages: a loose (very high recall but low precision) automatic matching followed by
manual checking to remove false positives (although some counts may still be slight over-estimates).
For Lerner’s (1996)opportunisticcases, we looked for filled pauses (‘er/erm’ etc.) or pauses explic-
itly annotated in the transcript (‘<pause>’), so counts in this case may be underestimates if short
pauses were not transcribed. We also chose some other broad categories based on our observations
of the most common cases. Results are shown in Table 4 (where the‖ token represents the split
point).9

person: Same- Cross-
all cross-turn (all)
N % N % N %

. . .‖ and/but/or . . . 748 39 306 42 116 36

. . .‖ so/whereas . . . 257 14 57 8 39 12

. . .‖ because . . . 77 4 32 4 3 1

. . . er/erm‖ . . . 35 2 21 3 12 4

. . .<pause> ‖ . . . 19 1 15 2 20 6

. . .‖ which/who/etc . . . 26 1 11 2 4 1

. . . instead of . . .‖ . . . 1 0 0 0 0 0

. . . said/thought/etc . . .‖ . . . 12 1 5 1 0 0

. . . if . . .‖ (then) . . . 18 1 10 1 2 1

. . . when . . .‖ (then) . . . 6 0 4 1 1 0
(other) 783 41 317 44 164 50
Total 1902 726 329

Table 4: Continuation categories

The most common of the CA categories can be seen to be Lerner (1996)’s hesitation-related
opportunisticcases, which make up 3-5% of same- and 10% of cross-person CCs, meaning cross-
person opportunistic cases are more common than same-person ones (same(cross-turn; 36 of 726)
vs other (32 of 329)χ2

(1) = 8.53, p = 0.00310). Interestingly, the breakdown of cases into those
where the antecedent ends with an unfilled pause versus those which endwith a filled pause also
shows a difference between same- and cross-person cases: an other person is more likely to offer
a continuation after an unfilled pause, than after a filled pause (antecedents ending in‘er(m)’ 35
continued by same, 12 by other; ending in ‘<pause>’ 19 continued by same, 20 by otherχ2

(1) =
6.05, p = 0.01). This finding backs up claims by Clark and Fox Tree (2002), that filled pauses
can be used to indicate that the current speaker’s turn is not yet finished and thus have the effect of
holding the floor.

Lerner’s compound TCU cases (instead of, said/thoughtetc,if-thenandwhen-then) account for
2-3% of same-person and 1% of cross-person CCs, though note that these could be underestimates,
as his non-syntactic cues (e.g. contrast stress and prefaced disagreements) could not be extracted.
Rühlemann’s (2007)sentence relativecases come next with over 1%.

9. Note that the categories in Table 4 are not all mutually exclusive (e.g. an example may have both an‘and’-initial
continuation and an antecedent ending in a pause), so column sums will not match totals shown.

10. For completeness, wherep > 0.001, we report exact probabilities but throughout adopt a criterion probability level
of < 0.05 for accepting or rejecting the null hypothesis.

293



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

In contrast, by far the most common pattern (for same- and cross-personCCs) is the addition of
an extending clause, either a conjunction introduced by‘and/but/or/nor’ (36-42%), or other clause
types with‘so/whereas/nevertheless/because’. There are differences in the proportions of the clause
types between same- and cross-person CCs, but further research and annotation is needed to confirm
whether this represents systematic differences in pragmatic use (as in Rühlemann’s (2007) sentence
relative study, where cross-person CCs more often expressed stance (speaker opinion) than same-
person CCs).

Split point Other less obviously categorisable cases make up 40-50% of continuations, in both
same- and cross- person cases, with the most common first words being‘you’, ‘it’ , ‘I’ , ‘the’, ‘in’
and ‘that’ . In terms of syntactic categories, manual examination of the data suggests that the split
point can occur at any point between words,11 even within what traditional theories of grammar
consider to be a single constituent,12 such as noun phrases and prepositional phrases (11, 12, 13,
14).

(11) D: Yeah I mean if you’re looking at quantitative things it’s really you know how much
actual- How much variation happens whereas qualitative is〈pause〉 you know what
the actual variations

U: entails

D: entails. you know what the actual quality of the variations are.[BNC G4V 114-117]

(12) M: We need to put your name down. Even if that wasn’t a

P: A proper conversation

M: a grunt. [BNC KDF 25-27]

(13) A: All the machinery was

G: [[All steam.]]

A: [[operated]] by steam [BNC H5G 177-179]

(14) K: I’ve got a scribble behind it, oh annual report I’d get that from.

S: Right.

K: And the total number of [[sixth form students in a division.]]

11. There is anecdotal evidence that CCs can also occur mid-word, aswhen someone completes a complex multi-syllabic
word for another person. Only one of our cross-person CCs occurred mid-word (shown in (i), from a doctor/patient
exchange), in which the whole word is also repeated, so we leave such considerations aside for now, though obviously
they have implications for e.g. the organisation of the lexicon.

(i) A: No it wasn’t Marvelon it was that Trin

D: Trin

A: Aye.

D: Trinordiol.

A: Mhm. [BNC G58 63-68]

12. Of course, different grammars may have different notions of constituency (such as thesurprising constituentsof CCG
(Steedman, 2000)) which these findings may have a bearing on, however, for the purposes of the current discussion,
we limit our notion of constituency to that of syntactic elements as in, for example transformational grammars, or
HPSG.

294



COMPOUND CONTRIBUTIONS

S: [[Sixth form students in a division.]] Right.

[BNC H5D 123-127]

To further test the finding that the split point can apparently occur between any types of words,
we annotated thecompletioncases for whether the split point occurred within a syntactic constituent,
or between constituents.13 For same-person cross-turn CCs, just over half are between-constituent
(111/213; 52%), whilst cross-person CCs appear to be more likely to occur within-constituent al-
though this trend is not significant (52/87; 60%;χ2

(1) = 3.49, p = 0.06). This finding appears
to be associated with repair (there seem to be more repairs in the within-constituent cases) but the
numbers are too small to be sure.

person: Same- Cross-
all cross-turn (all)
N % N % N %

Antecedent end-complete Y 136772 513 71 242 74
N 535 28 213 29 87 26

Continuation start-complete Y 22412 99 14 48 15
N 1678 88 627 86 281 85

Repair Y 77 4 34 5 32 10
N 1825 96 692 95 297 90

Total 1902 726 329

Table 5: Completeness and repair

Completeness Examination of theend-complete annotations shows that about 8% of s-units
in general are incomplete (930/11469), but that (perhaps surprisingly) only 64% (591/930) of these
get continued. This compares to 15% of end-complete s-units (1577/10539) that get continued
(χ2

(1) = 1315.90, p < 0.001), showing that although incomplete s-units are more likely to be
continued, incompleteness does not necessarily prompt the production ofa completion.

The majority of both same- and cross-person continuations (71% to 74%) continue an already
complete antecedent, with only 26-29% therefore beingcompletionsin the sense of e.g. Ono and
Thompson (1993). Interestingly, though, continuations are no more likely than other s-units to end
in a complete way themselves. In fact, continuations are significantly more likely than other s-units
to end in an incomplete way (273/2231 (12%) vs. 657/9238 (7%);χ2

(1) = 63.34, p < 0.001).
The frequent clausal categories from Table 4 are all much more likely to continue complete

antecedents than incomplete ones.14 This is not the case for the(other) category; again suggest-
ing that split points often occur at random points in a sentence, without regard to particular clausal
constructions. The continuations in the(other) category are far less likely to continue complete
antecedents than the easily classifyable categories from table 4 (220/481; 46% v. 535/574; 93%,
χ2
(1) = 289.76, p < 0.001).

13. Here we are concerned with only low-level syntactic constituency; wecounted a split point as within-constituent if
it fell within a noun phrase (e.g. between a determiner and noun), a prepositional phrase (e.g. between a preposition
and a noun phrase) or within a complex noun phrase (e.g. between an auxiliary and a head noun). Other cases (e.g.
between a verb and its object, or between clauses) were coded as between-constituent.

14. For the less frequent (e.g.‘if/then’, ‘instead of’) categories, the counts are too low to be sure.

295



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

Looking only at the general(other) category, we see that cross-person continuations more
often follow antecedents that end in a complete way than same-person continuations (89/164; 54% v.
131/317; 41%χ2

(1) = 7.30, p = 0.007). For both cross-person and same person cases, continuations
in the(other) category do not often start in a complete way (cross-person: 41/164; 25%, same-
person 94/317; 30%).

In general, however, continuations are more than twice as likely to start in a non-complete rather
than a complete way, even after complete antecedents.

Repair Explicit repair of the antecedent is not common, only occurring in just under 5% of CCs.
As might be expected, incomplete antecedents are more likely to be repaired (cross-turn (same and
cross-person); 51/300 17% vs. 15/755 2%,χ2

(1) = 82.51, p < 0.001). Cross-person continuations
are also significantly more likely to repair their antecedents than same-personcases (32/329; 10%
vs. 34/726; 5%,χ2

(1) = 9.82, p = 0.002).
In those CCs where the split point falls within a syntactic constituent, only 18% (18/102) of

same-person cases involve explicit repair at the start of the continuation,compared to 27% (14/52)
of cross-person CCs (the equivalent figures for CCs where the splitpoint is between constituents
are 12% (13/111) and 18% (6/35)). Although more data are required to see if these are genuine dif-
ferences, we know that repair in general is not common, so it appears that even when the split point
occurs mid-constituent, the participants generally are able to just go on extending the constituent
as if they were the original speaker. This might suggest that the parsing and generation mecha-
nisms are not required to back up to the beginning of a constituent in order toprocess or produce a
continuation (i.e. start with a new grammar rule). This seems to favour lexicalised or dependency-
based parsing models in that it suggests that the language processing mechanisms directly rely on
word-by-word dependencies rather than constituents/grammar rules.

Function of CCs We are concerned in this study primarily with the form, rather than the function,
of CCs. However, it is worth noting at this point that they can perform functions beyond merely
extending or completing an interlocutor’s contribution (see also Szczepek,2000b); and in some
cases are difficult to define functionally, and may even exhibit genuinemultifunctionality(see e.g.
Gregoromichelaki et al., 2009; Bunt, 2009). In (15), for example, J’scontinuation of M’s utterance
serves also as a request for confirmation:

(15) M: It’s generated with a handle and

J: Wound round?

M: Yes [BNC K69 109-112]

In many cases, the antecedent explicitly invites the hearer to complete the contribution, so that
antecedent and continuation form a question-answer pair, possibly withina single grammatical con-
stituent (16):

(16) J: The Holy Spirit is the one who gives us hope.

Mega.

I mean〈pause〉 this generation needs hope.

The Holy Spirit is one who〈pause〉 gives us?

U: Strength.

J: Strength.

296



COMPOUND CONTRIBUTIONS

Yes, indeed.

〈pause〉 The Holy Spirit is one who gives us?〈pause〉

U: Comfort.

Yes. [BNC HDD 274-283]

This phenomenon can also happen in cases of clarification-request/clarification-reply pairs (see
Purver et al.’s (2003)gapcategory), e.g. (17):

(17) G: Cos they〈unclear〉 they used to come in here for water and bunkers you see.

A: Water and?

G: Bunkers, coal, they all coal furnace you see,〈clears throat〉 and we er they’d come in
and we used to fill them up with coal, whatever they wanted〈cough〉 lot of that went
over the side〈unclear〉 coal, beautiful coal that was. [BNC H5H 59-61]

With the range of possibilities regarding where the split point is able to occur,including po-
tentially within a word (see footnote 11) it is hard to see how compound contributions could be
characterised as a well-defined syntactic phenomenon, a separate grammatical fragment category,
or a sub-class ofnon-sentential utterance(Ferńandez and Ginzburg, 2002). Moreover, there seems
no reason to associate either antecedent or continuation with particular semantic categories or spe-
cific pragmatic speech-act information, as they seem to serve a wide rangeof purposes in dialogue:
from assisting a speaker with lexical access, to eliciting a response to a query, to covertly offering a
suggestion or asking a clarification.

Summary The results here show that CCs are common in dialogue. Split points may be possible
at any syntactic point, but there appear to be (possibly pragmatic) constraints on where they are
likely to appear: they are far more likely after complete antecedents, although relatively few of them
occur in the highly projectable positions studied by e.g. Lerner.

There are interesting differences between same-person CCs and cross-person CCs; firstly, same-
person CCs are over twice as common as cross-person. Cross-person continuations are more likely
to start with explicit repair/reformulation of the antecedent; this might be considered surprising, as
self-repair is preferred in general (Schegloff et al., 1977) althoughwe have no comparable figures
for repair at other points in the turn. However, it is interesting to note that a CC, in virtue of being
constructed as a continuation of the speakers utterance, may provide a device that enables a less
exposed form of other repair.

Outside the frequent clausal or CA categories, cross-person CCs are also more likely than same-
person to continue a complete antecedent; and they are more likely where the antecedent ends in
an unfilled pause rather than a filled one. This suggests an effect on turn-taking expectations, and
that continuations may be systematically invited by a speaker or designed as though they are natural
continuations of contributions that could be treated as complete. We will returnto these points in
the general discussion.

5. Study 2: Experimental Manipulation

While the corpus study of Section 4 provides us with useful information concerning the nature and
frequency of CCs and their various sub-categories, it can tell us nothing about the effect of CCs on
the dynamics of a conversation.

297



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

From a processing point of view, we might intuitively predict that cross-person CCs ought to
be more difficult for a third party to process than same-person CCs, as information from potentially
conflicting sources must be integrated and interpreted as a single syntactic unit. Conversely, some
models (e.g. Dynamic Syntax, Cann et al. (2005)) would predict that thereshould be no additional
processing costs.

The corpus study also suggests pragmatic effects associated with CCs. Are cross-person CCs
indicative of particularly close coordination (and thus of Schegloff’s ‘parties’ as Lerner suggests),
which might facilitate understanding, or are they viewed as impolite which may addadditional
implications and disrupt the flow of the conversation?

The experiment reported here is, we believe, the first controlled manipulation of compound
contributions during an unfolding interaction. The allows us to directly compare the effects of
same-person and cross-person CCs on participants in a dialogue.

The effects of seeing a CC on a dialogue in progress were tested using theDialogue Experimen-
tation Toolkit (DiET) chat tool, which enables text dialogues to be experimentally manipulated (see
Healey et al., 2003).

Of course, text based chat is different to face-to-face dialogue in several ways, and while clearly
an interesting field of study in itself (Rosé et al. (2003), for example, compare text and speech based
tutoring systems), there are important questions as to whether the results from our corpus study
are generalisable to such a different modality. The most obvious differences are attributable to the
channel of communication; speech versus text. In text-based chat suchas MSN Messenger and the
chat tool reported here, participants compose their turns in private before sending them to the other
participants. This means that they can revise or even delete their turns without their interlocutors
being aware of the revisions, unlike in face-to-face dialogue where overt repairs are necessarily
shared. It also means that participants can compose their next turns simultaneously, meaning that
the linearity of turn-taking in dialogue is lost. Linked to this is the fact that, unlike inface-to-face
dialogue, participants engaged in a text chat are not typically co-present. Although this means that
a number of non-linguistic cues are unavailable, this is also true in telephone conversations, for
example, so should not be taken as a reason for rejecting the dialogic nature of text chat.

Despite these differences, there are also important similarities between text chat and face-to-face
dialogue. Both involve the use of interlocutors’ language resources to communicate, and text chat
also exhibits many features which are generally seen in spoken dialogue, but not in either spoken
monologue or written text. These include the use of non-sentential utterances such as clarification
requests (Purver et al., 2003) and acknowledgements (Fernández and Ginzburg, 2002). Importantly
for the study reported here, CCs also occur naturally in text-based chat(see, for example, (7) and
(18), taken from the DiET chat tool environment).

(18) U: i agree tom needs to be there

A: but one of them has to go to save the other 2

R: and what about the cancer research plan ??

According to a preliminary corpus study (Eshghi, 2009) CCs occur as frequently in text chat
as they do in face-to-face dialogue. In a total of 2377 text contributions, there were 493 CCs,
of which 112 were cross-person and 381 were same-person CCs. Overall this proportion of CCs
is not different to that from our BNC corpus study. In the text chat corpus there was a higher
proportion of cross-person CCs than expected from our BNC results (112 out of 2377 versus 326

298



COMPOUND CONTRIBUTIONS

out of 11469χ2
(1) = 22.46, p < 0.001) which could be related to the task-based nature of the text

chat (the dialogues analysed in Eshghi (2009) are three-way tangramtask conversations between
two directors and a matcher), or possibly due to the way turns are transcribed into consecutive
contributions in the BNC, as previously discussed.

5.1 Method

In the DiET chat tool, interventions can be introduced into a dialogue in real time, thus causing a
minimum of disruption to the natural ‘flow’ of the conversation. In this experiment, a number of
genuine single contributions in a text-based three-way conversation wereartificially split into two
parts. In some conditions, both parts still appeared to originate from the genuine source (“speaker”),
thus appearing as a same-person CC. In other conditions, one or both parts seemed to come from
another participant, thus appearing either as an cross-person CC, or as a same-person CC generated
by the “wrong” person.

5.1.1 MATERIALS

The Balloon Task Theballoon taskis an ethical dilemma requiring agreement on which of three
passengers should be thrown out of a hot air balloon that will crash, killing all the passengers, if one
is not sacrificed. The choice is between a scientist, who believes he is on thebrink of discovering
a cure for cancer, a woman who is 7 months pregnant, and her husband,the pilot. This task was
chosen on the basis that it should stimulate discussion, leading to dialogues ofa sufficient length to
enable an adequate number of interventions.

The DiET Chat Tool The DiET chat tool itself is a custom built Java application consisting of
two main components: user interface and server console.

User interface The user interface is designed to look and feel like common instant messaging
applications e.g. Microsoft Messenger. It consists of a display split into two windows, separated by
a status bar, which indicates whether any other participant(s) are activelytyping (see Figure 1). The
ongoing dialogue, consisting of both the nickname of the contributor and theirtransmitted text, is
shown in the upper window. In the lower window, participants type and revise their contributions,
before sending them to their co-participants. All key presses are time-stamped and stored by the
server.

Server Console All text entered is passed to the server, from where it is relayed to the other
participants. No turns are transmitted directly between participants. Prior to being relayed, some
turns are altered by the server to create fake CCs.

This is carried out automatically. A genuine single-person contribution is splitaround a space
character near the centre of the string. The part of the turn before the space is relayed first, as the
antecedent, followed by a short delay during which no other turns may be sent. This is followed
by the continuation (the part of the turn after the space), as if they were in fact two quite separate,
consecutive contributions. In every case, the server produces two variants of the compound con-
tribution, relaying different information to both recipients. Each time an intervention is triggered,
one of the two recipients receives a same-person CC from theactual source of the contribution
(henceforth referred to as anAA-split). The other recipient receives one of three, more substantial,
manipulations: a same-person CC that wrongly attributes both antecedent and continuation to the

299



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

Figure 1: The user interface chat window (as viewed by participant ‘sam’)

other recipient (aBB-split); a cross-person CC whose antecedent comes from from the actual origin
and continuation from the other recipient (anAB-split), or vice-versa (aBA-split).

This allows us to create a2×2 factorial design which separates potential effects of ‘floor change’
i.e. whether the original speaker finishes the CC or another participant appears to, from effects of
‘same/other’ i.e. whether a the two halves of the CC or appear to be produced by the same speaker
or by two different speakers. This contrast is shown in table 6.

A types:

Should we start now
B sees (AA intervention):

A: Should we
A: start now
C sees (one of):

AB intervention: BA intervention: BB intervention:

A: Should we B: Should we B: Should we
B: start now A: start now B: start now

Table 6: Comparison of split types

The intervention is triggered every 10 turns, and restricted such that the participant who receives
the non AA-split is rotated (to ensure that each participant only sees any of the more substantially
manipulated interventions every 30 turns). Which of the three non AA-splits they see (AB, BA or
BB) is, however, generated randomly.

300



COMPOUND CONTRIBUTIONS

5.1.2 SUBJECTS

41 male and 19 female native English speaking undergraduate students were recruited for the ex-
periment, in groups of three to ensure that they were familiar with each other.All had previous
experience of internet chat software such as Microsoft Messengerand each was paid£7.00 for their
participation.

5.1.3 PROCEDURE

Each of the triad of subjects was sat in front of a desktop computer in separate rooms, so that they
were unable to see or hear each other. Subjects were asked to follow the on-screen instructions, and
input their e-mail address and their username (the nickname that would identify their contributions
in the chat window). When they had entered these, a blank chat window appeared, and they were
given a sheet of paper with the task description. Participants were instructed to read this carefully,
and begin discussing the task with their colleagues via the chat window once they had done so. They
were told that the experiment was investigating the differences in communicationwhen conducted
using a text-only interface as opposed to face-to-face. Additionally, subjects were informed that the
experiment would last approximately 20-30 minutes, and that all turns would be recorded anony-
mously for later analysis. Once all three participants had been logged on, the experimenter went to
sit at the server machine, a fourth desktop PC out of sight of all three subjects, and made no further
contact until after at least 20 minutes of dialogue.

5.1.4 ANALYSIS

As production and receipt of contributions sometimes occurs in overlap in text chat, it is not possible
to say definitively when one contribution is made in direct response to another.15 We therefore chose
to measure all the contributions produced by both recipients between the mostrecent intervention
and the next intervention, averaged to produce one data point per recipient per intervention. This
means that there are two data points for each intervention (one for each ofthe participants who saw
a fake compound contribution).

The data were analysed according to two factors in a2×2 factorial design;same/other– whether
both parts of the compound contribution appeared to come from the same-person, or from different
sources ([AA and BB] vs [AB and BA]), andfloor change– whether the continuation part of the CC
appeared to come from the genuine source or the other participant ([AA and BA] vs [AB and BB]),
with participant as a random factor.

Measures selected for analysis weretyping time of turn(the time, in milliseconds, between the
first key press in a turn and sending the turn to the other participants by hittingthe return key) and
length of turn in charactersas measures of production;deletes per character(the number of keyed
deletes divided by the total number of characters) as a measure of revisions; andtyping time per
characteras a measure of speed. Data in tables are displayed in the original scale of measurement.
However, as inspection of the data showed that they were not normally distributed, logarithmic

15. In online chat, participants can compose their next contributions simultaneously, and contributions under construction
when another is received can be subsequently revised, prior to transmission. This means that a genuine response
to a compound contribution might have a negative start time. However, theinclusion of cases where the whole
contributions was constructed after receiving the CC (an arbitrary cut-off point, which would catch some contributions
that were responses to earlier contributions in the dialogue, and miss somewhich were begun before the intervention
was received and subsequently revised) should impose the same levelof noise in all cases.

301



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

transformations (usingloge) were applied to the typing time of turn and length of turn in characters
measures prior to all inferential statistical analyses, resulting in data distributions that were not
significantly different from a normal distribution (using Shapiro-Wilk tests:typing time of turn
W = 0.998, p = 0.882; length of turn in charactersW = 0.995, p = 0.100). For the proportional
measures of deletes per character and typing time of character, which violate normality assumptions
even after transformations, alternative analyses were used.

The Generalized Linear Model (GZLM) extends the General Linear Model (GLM; which in-
cludes ANOVAs and linear regression models) to include response variables that follow any expo-
nential probability distribution, including e.g. poisson, binomial and gamma distributions. GZLMs
use maximum likelihood estimation to fit the model to the data (and provide parameter estimates).
Generalized Estimating Equations (GEE) extend GZLM further by allowing fornon-independent
data, such as repeated measures and clustered data. Using a GEE analysis (see Liang and Zeger,
1986; Ballinger, 2004) on these variables therefore allows for both the non-normality of the data,
and within-subject correlations.

5.2 Results

A post-experimental questionnaire and debriefing showed that, with the exception of one subject,
who had taken part in a previous chat tool experiment and was therefore aware that manipulations
may occur, none of the participants were aware of any interventions.

Of the 253 interventions to which at least one recipient responded, 89 were AA/AB splits, 99
were AA/BA splits and 65 AA/BB splits. This means there were 506 potential responses, however,
in 16 cases, only one of the recipients produced a response, leaving 490 data points. Table 7 shows
the actual n values in each case.

Typing time / turn (ms) Typing time / char (ms)
Condition Mean (s.d.) Mean (s.d.) N

AA 11122.27 (14413.5) 475.45 (558.92) 246
AB 12500.98 (10944.6) 523.56 (1036.00) 89
BA 9800.77 (8810.3) 357.76 (316.15) 92
BB 11561.67 (10138.4) 479.51 (396.07) 63

Table 7: Typing time of turn and typing time per character by type of intervention

2 × 2 ANOVAs (with participant as a random effect)16 show a significant main effect of floor
change17 on the log transformed typing time of turn (see table 7), with participants taking longer
over their turns in the AB and BB conditions (F(1,288) = 6.563, p = 0.012). There was no main
effect of same/other (F(1,288) = 0.001, p = 0.980), and no effect of interaction (F(1,288) = 1.259,
p = 0.270), though there was a main effect of participant (F(59,288) = 4.565, p = 0.008) showing
that there was high individual variation for this measure.

16. We account for between subject variation by including subject as a random factor, meaning that there is more than
one datapoint per subject (and, in effect, a2× 2× 60 model). There are 490 datapoints between 60 subjects. As we
carried out a full factorial model, the numerator (error) degrees offreedom that resulted from this model was 288.

17. A significant effect is one in which the p-value, the probability of the observed data being sampled if the null hypoth-
esis were true, is below some criterion value. We adopt the criteria of significance as lower than 5% probability.

302



COMPOUND CONTRIBUTIONS

There were no significant effects on length of turn in characters (same/other,F(1,288) = 1.709,
p = 0.194, floor changeF(1,288) = 0.194, p = 0.341).

A 2×2 GEE model with participant as a subject effect (using the gamma distribution, goodness
of fit quasi log-likelihood (QIC) = 149.482)18 showed a marginally significant main effect of floor
change on typing time per character (Model effect; Wald-χ2 = 3.820, p = 0.051. Parameter
estimate;B = −0.281, Wald-χ2 = 7.192, p = 0.007) and a main effect of participant (Wald-
χ2 = 258468, p < 0.001). There was no main effect of same/other, and no interaction effects.

For deletes per character, a2 × 2 GEE model with participant as a subject effect (using the
negative binomial distribution,19 goodness of fit quasi log-likelihood (QIC) = 566.574) showed a
significant main effect of same/other (Model effect; Wald-χ2 = 9.617, p = 0.002. Parameter
estimate;B = −0.492, Wald-χ2 = 12.226, p < 0.001) and a main effect of participant (Wald-
χ2 = 487986, p < 0.001). There was no main effect of floor change, and no interaction effects.

Condition Mean (s.d.)
(ms/char)

AA 0.108 (0.16)
AB 0.094 (0.13)
BA 0.071 (0.10)
BB 0.138 (0.17)

Table 8: Deletes per character by type of intervention

As the experiment was looking for generic effects of CCs on the dialogue,the location of the
split points was arbitrary. In order to test for effects of split point, post-hoc analyses were carried out
to ascertain whether other observed contrasts in the corpus had any effects on processing of apparent
CCs. The fake CCs were coded according to three factors; standalonecoherence (as judged by the
authors) of the antecedent and continuation20 (see table 9) and whether the split point fell within
or between a syntactic constituent. There were no effects of first or second half coherence on any
of the variables, and no interaction effects. There were also no main effects of whether the split
point fell within or between a constituent; (log transformed) typing time of turn (F(1,204) = 0.262,
p = 0.435); (log transformed) number of characters (F(1,204) = 1.760, p = 0.189); typing time per
character (Wald-χ2 = 0.550, p = 0.458) deletes per character (Wald-χ2 = 0.285, p = .594) and
no interaction effects with same/other or floor change. These results are consistent with the finding
from the corpus that the split point may be able to occur anywhere syntactically, though the lack
of any observed effects could be due to low power caused by the relatively small numbers of some
groups.

18. The model distributions were chosen on the basis of being the best fitto the data, as indicated by the lowest quasi
log-likelihood score.

19. Each key press can be seen as a delete or not-a-delete.
20. These judgements are simply a yes/no answer to the question ‘could thiscontribution be interpreted as complete in

its own right?’, i.e. analogous to theend-completeandstart-completeannotation tags in the corpus study, such that
an antecedent (first part) judged to be able to stand alone can be considered end-complete and a continuation (second
part) judged to be able to standalone can be considered start-complete. The difference in tagging conventions is due
to the fact that in the chat tool environment turns can be revised prior to sending, and therefore might be considered
to be a unit by its sender, even if the fractured nature of text chat meansthat it might not constitute a syntactically
complete sentence.

303



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

Part of CC Coherent
N

First Second 1st 2nd
what the hell is that Y N 131
the woman is pregnant she should stay Y Y 52
these people said you did something N Y 43
I think this is also the wish of the doctor N N 264

Table 9: Examples of standalone coherence judgement examples

5.3 Discussion

Given the novelty of the method and the lack of other experimental studies of CC’s to cross-check
against, the results of this experiment must be interpreted with caution. Nonetheless, we believe
that the results summarised in table 10, below, do bear on the questions raisedin section 3.2.

Effect of
Dependent variable
and direction of effect

Floor Change Typing Time (per turn and char)
(AB ∧BB) > (AA ∧BA)

Same/Other Deletes
(AA ∧BB) > (AB ∧BA)

Table 10: Summary of significant effects

Firstly, it is important to note that introducing fake CCs did have measurable effects on the
ongoing dialogue, despite participants being unaware of either the intervention, or their effects.
This in itself might be seen as surprising, as if the intervention were highly disruptive, we would
presumably expect subjects to notice it.

Though typing time is a fairly crude measure21 one possible explanation for participants taking
longer over the production of a turn (independently of length of turn in characters) could be due to
problems arising in the local organisation of turn-taking (Sacks et al., 1974). A participant who has
seen a floor change intervention (Participant C) may be taking longer overtheir turns because there
is less pressure on them to take a turn. C will falsely believe that the fake source (Participant B) has
just completed a turn, and will therefore not expect them to take the floor. Additionally, the genuine
source (Participant A) will not be taking the floor because they have justcompleted a turn (though C
does not know this). However, this effect of floor change could also be due to the confounding fact
that when one of the recipients sees a floor change CC, and the other recipient (as always) sees an
AA-split, the two are left with different impressions about who made the finalcontribution (i.e. the
continuation part of the fake CC) and thus have potentially conflicting expectations regarding who
is entitled to speak next. Whether or not these explanations are correct, theeffect does suggest that
at some level participants are sensitive to specific interlocutors – note that the difference cannot be

21. For example, the additional typing time may fall at the end of a turn (before pressing enter) suggesting that participants
are reviewing their responses more carefully before sending them, orit may be a general effect spread evenly across
the turn.

304



COMPOUND CONTRIBUTIONS

simply attributable to a mismatch between who appears to be speaking and what sort of thing they
would say because then we would expect turns following the BA interventionto be equally affected.

Independently of a change of floor, seeing a CC that appears to be shared between speakers
also has an impact on the conversation, seen in the amount of revision undertaken in formulating
responses (deletes). Perhaps surprisingly, in this case, participants who have seen a CC that was
apparently co-constructed by both their interlocutors revise their turns less than after a same-person
CC. One reason why participants might worry less about precisely formulating their turns following
a cross-person CC is that it could have the effect on the recipient of suggesting that the two other
participants are highly coordinated. One possible interpretation of this couldbe that they have
formed a ‘party’ (Schegloff, 1995) with respect to the decision of who tothrow out of the balloon.
This might be understood as signalling the formation of a strong coalition between the other two
participants, making the recipient behave as though they are resigned to thedecision of this coalition.
(19), taken from the transcripts shows an example where this appears to be the case (the ‘fake’ part
of the CC is shown in bold).

(19) AB-Split showing apparent coalition between ‘B’ and ‘D’

B: and he can tell his formula

D: to tom and susie

Note that this is not the same as the effect on the typing time of turn, whereby participants are
less rushed when seeing a change of floor. Deletes, in contrast, indicatehow carefully participants
are constructing their turns.

6. General Discussion and Conclusions

As discussed in the introduction, CCs are of interest for many different groups of researchers. Our
corpus study shows that nearly one fifth of all contributions in naturally occurring dialogue continue
some previous contribution, indicating the scale of the phenomenon. Just thesub-set of cross-person
CCs accounts for 3% of all dialogue contributions, comparable to the frequency of clarification
requests (see Purver et al., 2003; Rodrı́guez and Schlangen, 2004), widely studied by dialogue
theorists (e.g. Ginzburg and Cooper, 2004).

Although most of the categories of CC described by conversation analystsappear, these cate-
gories do not correspond to the most frequent in the BNC, and do not adequately characterise all the
CCs observed in the present analysis. The corpus results show no evidence that syntax places sig-
nificant constraints on where a split point can occur and the experimentalresults are consistent with
this. Participants were able to process and interpret fake CCs successfully despite their arbitrary
split points and were also not explicitly aware of the experimental manipulation.This is consistent
with models that advocate highly coordinated resources between interlocutors and, moreover, the
need for incremental means of processing that operate on at least a word-by-word basis (Purver
et al., 2006; Skantze and Schlangen, 2009).

However, that CCs may be able to occur anywhere in a syntactic sequenceis not to say that they
necessarily or usually do. Both the corpus study, in which cross-person CCs occur more frequently
after an unfilled (rather than a filled) pause and more often follow an end-complete antecedent
than same-person CCs (in the not obviously classifiable cases), and the experiment, in which con-
founding expectations lead to additional response time, suggest that conversational expectations (in-

305



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

cluding for example around turn-taking) play some role above and beyondgrammatical/linguistic
resources.

All CCs Other person CCs
BNC Experiment BNC Experiment

Y 160972% 183 37% 24274% 69 38%
N 62228% 307 63% 8726% 112 62%
Total 2231 490 329 181

Table 11: Antecedent end-completeness: Comparison of the distribution ofactual (corpus) and ar-
bitrary (experiment) split points

These considerations are backed up by the data shown in Table 11. This table shows the distribu-
tion of antecedent end-completeness in the annotated corpus, compared tothe distribution obtained
in the experiment. As can be clearly seen, when strings are artificially split in aarbitrary fashion,
the ‘antecedent’ is far less likely to end in a complete way than actually occurs inthe genuine CCs
in the corpus. This suggests that continuations are systematically designed (and their split points
chosen) as extensions of contributions that could be treated as already complete.

In terms of the effects that CCs have on the ongoing dialogue, the experiment suggests that Ler-
ner’s hypothesis that cross-person CCs might demonstrate party membership may well be correct;
it also clearly demonstrates that the participants in a dialogue are not interchangeable. These results
do not, of course, prejudice the claim that, at a purely mechanistic level, people could anticipate
the structures needed to complete a turn (as the interactive alignment model suggests); they do not
tell us about the actual production of compound contributions, but ratherabout the effect they have
on the conversation, so do not provide unequivocal evidence in support of one theory over another.
They do however indicate that if we wish to treat a jointly produced CC as signalling especially
strong alignment, then we need to examine factors other than simply syntax.

They also offer some interesting pointers for further research: if parties are genuine conver-
sational entities, then we might expect dialogue phenomena to have distributional patterns which
reflect this. Consistent with our corpus results as regards other-person repair, for example, is the
speculative hypothesis that people might be structuring continuations precisely as the preferred
‘self’-repairs, where ‘self’ can be taken to mean within-party.

From a computational modelling point of view, there is some good news: as start-completeness
of continuations is rare, a dialogue system may have a chance of detecting continuations from
surface characteristics of the input (though note that we did not investigate the general prevalence
of start-incomplete s-units in the corpus). There is bad news too, though: as theredo not seem to be
strict syntactic restrictions on where the split point can occur, there may beno grammatical features
that can be reliably employed to this end. In addition, antecedents do not endin an incomplete
way as commonly as might be expected, and long distances between antecedent and continuation
are possible. Detecting continuations and locating their antecedents is therefore unlikely to be a
straightforward task for automated systems.

306



COMPOUND CONTRIBUTIONS

6.1 Further Work

Split point For implementational purposes, additional corpus analysis needs to be carried out
regarding the distribution of the different syntactic points at which CCs canand do appear; further
experiments are also planned which examine the effect on processing of manipulating the syntactic
position of the split point (for example, inserting splits before or after determiners).

Continuation form Further corpus analysis is also required to investigate any systematic differ-
ences in theform of continuations, and the interaction of this with the properties of the antecedent
and conversational genre (including the BNC’s “context-governed” or “demographic” face-to-face
dialogues and text-based chat).

Ownership A further interesting question regards who can be said to take responsibilityfor, or
‘own’ a jointly produced CC. Insights into this might come from Lerner (2004), which discusses
CCs that occur incollaborative turn sequences: if the original speaker maintains authority over the
content of the collaboratively constructed contribution, they may ratify another speaker’s continua-
tion (for example by repeating or acknowledging it), but may alternatively strategically offer their
own (delayed completion,Lerner, 1996). Party-membership, or otherwise, may be influenced by
such additional conversational contributions.

Character-by-character experiments Due to the design of the experiment, the floor change ef-
fects might, as discussed, be because in floor change cases the two recipients will have been left
with the impression that a different person made the final contribution. This means that there may
be a an effect of confounded listener expectation (though see Schober and Brennan, 2003, for dis-
cussion) – though note that this does not have any bearing on the observed differences on deletes
after a cross-person CC. Because of this, and the already noted potential problems of linearity in
text-based chat, a follow-up study using a character-by-character chat tool interface is already under
way. This more directly enforces turn-taking, as it does not allow participants to formulate their
turn before communicating it; each character is transmitted as and when it is entered.

References

G.A. Ballinger. Using generalized estimating equations for longitudinal data analysis. Organiza-
tional Research Methods, 7(2):127–150, 2004.

Harry Bunt. Multifunctionality and multidimensional dialogue semantics. InProceedings of Dia-
Holmia, 13th SEMDIAL Workshop, 2009.

Lou Burnard. Reference Guide for the British National Corpus (World Edition). Oxford Uni-
versity Computing Serviceshttp://www.natcorp.ox.ac.uk/docs/userManual/,
2000. URLhttp://www.natcorp.ox.ac.uk/docs/userManual/.

Okko Buß, Timo Baumann, and David Schlangen. Collaborating on utterances with a spoken di-
alogue system using an ISU-based approach to incremental dialogue management. InProceed-
ings of the SIGDIAL 2010 Conference, pages 233–236, Tokyo, Japan, September 2010. Asso-
ciation for Computational Linguistics. URLhttp://www.sigdial.org/workshops/
workshop11/proc/pdf/SIGDIAL42.pdf.

307



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

Ronnie Cann, Ruth Kempson, and Lutz Marten.The Dynamics of Language. Elsevier, Oxford,
2005.

Jean Carletta. Assessing agreement on classification tasks: The kappa statistic. Computational
Linguistics, 22(2):249–255, 1996.

Herbert H. Clark.Using Language. Cambridge University Press, 1996.

Herbert H. Clark and Jean E. Fox Tree. Usinguh andum in spontaneous speaking.Cognition, 84
(1):73–111, 2002.

Jan Peter de Ruiter and Mattia van Dienst. Completing other people’s utterances: evidence for
forward modeling in conversation. ms., in preparation.

David DeVault, Kenji Sagae, and David Traum. Can I finish? learning whento respond to incremen-
tal interpretation results in interactive dialogue. InProceedings of the SIGDIAL 2009 Conference,
pages 11–20, London, UK, September 2009. Association for Computational Linguistics. URL
http://www.aclweb.org/anthology/W/W09/W09-3902.

Arash Eshghi.Uncommon ground: The distribution of dialogue contexts. PhD thesis, School of
Electronic Engineering and Computer Science, Queen Mary University ofLondon, 2009.

Raquel Ferńandez and Jonathan Ginzburg. Non-sentential utterances: A corpus-based study.Traite-
ment Automatique des Langues, 43(2), 2002.

Victor Ferreira. Is it better to give than to donate? Syntactic flexibility in language production.
Journal of Memory and Language, 35:724–755, 1996.

Barbara A. Fox. Principles shaping grammatical practices: an exploration.Discourse Studies, 9(3):
299, 2007.

Andrew Gargett, Eleni Gregoromichelaki, Ruth Kempson, Matthew Purver,and Yo Sato. Grammar
resources for modelling dialogue dynamically.Cognitive Neurodynamics, 3(4):347–363, 2009.
ISSN 1871-4080. URLhttp://dx.doi.org/10.1007/s11571-009-9088-y.

Jonathan Ginzburg and Robin Cooper. Clarification, ellipsis, and the nature of contextual updates
in dialogue.Linguistics and Philosophy, 27(3):297–365, 2004.

Eleni Gregoromichelaki, Yo Sato, Ruth Kempson, Andrew Gargett, and Christine Howes. Dialogue
modelling and the remit of core grammar. InProceedings of IWCS, 2009.

Eleni Gregoromichelaki, Ruth Kempson, Matthew Purver, Greg J. Mills, Ronnie Cann, Wilfried
Meyer-Viol, and Pat Healey. Incrementality and intention-recognition in utterance processing.
Dialogue and Discourse, to appear, 2011. to appear.

Markus Guhe.Incremental Conceptualization for Language Production. NJ: Lawrence Erlbaum
Associates, 2007.

Makoto Hayashi. Where Grammar and Interaction Meet: A Study of Co-Participant Completion in
Japanese Conversation.Human Studies, 22(2):475–499, 1999.

308



COMPOUND CONTRIBUTIONS

Patrick Healey, Matthew Purver, James King, Jonathan Ginzburg, and Greg Mills. Experimenting
with clarification in dialogue. InProceedings of the 25th Annual Meeting of the Cognitive Science
Society, Boston, Massachusetts, August 2003.

Marja-Liisa Helasvuo. Shared syntax: the grammar of co-constructions.Journal of Pragmatics, 36
(8):1315–1336, 2004.

Ruth Kempson, Wilfried Meyer-Viol, and Dov Gabbay.Dynamic Syntax: The Flow of Language
Understanding. Blackwell, 2001.

Gene H. Lerner. On the syntax of sentences-in-progress.Language in Society, pages 441–458,
1991.

Gene H. Lerner. On the “semi-permeable” character of grammatical units in conversation: Condi-
tional entry into the turn space of another speaker. In Elinor Ochs, Emanuel A. Schegloff, and
Sandra A. Thompson, editors,Interaction and grammar, pages 238–276. Cambridge University
Press, 1996.

Gene H. Lerner. Collaborative turn sequences. InConversation analysis: Studies from the first
generation, pages 225–256. John Benjamins, 2004.

Gene H. Lerner and Tomoyo Takagi. On the place of linguistic resources inthe organization of
talk-in-interaction: A co-investigation of English and Japanese grammatical practices.Journal of
Pragmatics, 31(1):49–75, 1999.

Kung-Yee Liang and Scott L. Zeger. Longitudinal data analysis using generalized linear models.
Biometrika, 73(1):13, 1986. ISSN 0006-3444.

Tsuyoshi Ono and Sandra A. Thompson. What can conversation tell usabout syntax. In P.W. Davis,
editor,Alternative Linguistics: Descriptive and Theoretical Modes. Benjamin, 1993.

Martin Pickering and Simon Garrod. Toward a mechanistic psychology of dialogue. Behavioral
and Brain Sciences, 27:169–226, 2004.

Massimo Poesio and Hannes Rieser. Completions, coordination, and alignment in dialogue. Di-
alogue and Discourse, 1:1–89, 2010. ISSN 2152-9620. URLhttp://elanguage.net/
journals/index.php/dad/article/view/91/512.

Matthew Purver, Jonathan Ginzburg, and Patrick Healey. On the means for clarification in dialogue.
In R. Smith and J. van Kuppevelt, editors,Current and New Directions in Discourse & Dialogue,
pages 235–255. Kluwer Academic Publishers, 2003.

Matthew Purver, Ronnie Cann, and Ruth Kempson. Grammars as parsers:Meeting the dialogue
challenge.Research on Language and Computation, 4(2-3):289–326, 2006.

Kepa Rodŕıguez and David Schlangen. Form, intonation and function of clarification requests in
German task-oriented spoken dialogues. InProceedings of the 8th Workshop on the Semantics
and Pragmatics of Dialogue (SEMDIAL), Barcelona, Spain, July 2004.

309



HOWES, PURVER, HEALEY, M ILLS AND GREGOROMICHELAKI

Carolyn P. Rośe, Diane Litman, Dumisizwe Bhembe, Kate Forbes, Scott Silliman, Ramesh Sri-
vastava, and Kurt VanLehn. A comparison of tutor and student behavior in speech versus text
based tutoring. InProceedings of the HLT-NAACL 03 workshop on Building educational ap-
plications using natural language processing-Volume 2, page 37. Association for Computational
Linguistics, 2003.

Christoph R̈uhlemann.Conversation in context: a corpus-driven approach. Continuum, 2007.

Harvey Sacks.Lectures on Conversation. Blackwell, 1992.

Harvey Sacks, Emanuel A. Schegloff, and Gail Jefferson. A simplestsystematics for the organiza-
tion of turn-taking for conversation.Language, 50(4):696–735, 1974.

Emanuel A. Schegloff. Parties and talking together: Two ways in which numbers are significant
for talk-in-interaction. Situated order: Studies in the social organization of talk and embodied
activities, pages 31–42, 1995.

Emanuel A. Schegloff. Turn organization: One intersection of grammar and interaction. In Elinor
Ochs, Emanuel A. Schegloff, and Sandra A. Thompson, editors,Interaction and grammar, pages
52–133. Cambridge University Press, 1996.

Emanuel A. Schegloff, Gail Jefferson, and Harvey Sacks. The preference for self-correction in the
organization of repair in conversation.Language, 53(2):361–382, 1977.

David Schlangen. From reaction to prediction: Experiments with computationalmodels of turn-
taking. In Proceedings of the 9th International Conference on Spoken LanguageProcessing
(INTERSPEECH - ICSLP), Pittsburgh, PA, September 2006.

Michael F. Schober and Susan E. Brennan. Processes of interactive spoken discourse: The role of
the partner.Handbook of discourse processes, pages 123–64, 2003.

Gabriel Skantze and David Schlangen. Incremental dialogue processing in a micro-domain. In
Proceedings of the 12th Conference of the European Chapter of the ACL(EACL 2009), pages
745–753, Athens, Greece, March 2009. Association for ComputationalLinguistics. URLhttp:
//www.aclweb.org/anthology/E09-1085.

Kristina Skuplik. Satzkooperationen. definition und empirische untersuchung. SFB 360 1999/03,
Bielefeld University, 1999.

Mark Steedman.The Syntactic Process. MIT Press, Cambridge, MA, 2000.

Beatrice Szczepek. Formal Aspects of Collaborative Productions in English Conversation. In-
teraction and Linguistic Structures (InLiSt),http://www.uni-potsdam.de/u/inlist/
issues/17/, 2000a.

Beatrice Szczepek. Functional Aspects of Collaborative Productions inEnglish Conversation.In-
teraction and Linguistic Structures (InLiSt),http://www.uni-potsdam.de/u/inlist/
issues/21/, 2000b.

Hongyin Tao and Michael J. McCarthy. Understanding non-restrictivewhich-clauses in spoken
English, which is not an easy thing.Language Sciences, 23(6):651–677, 2001.

310