Journal of Machine Learning Research-- Microsoft Word Template
Dialogue and Discourse 8(2) (2017) 149-166 doi: 10.5087/dad.2017.207
©2017 Ludivine Crible and Maria Josep Cuenca
This is an open-access article distributed under the terms of a Creative Commons Attribution License
(http://creativecommons.org/licenses/by/3.0/).
Discourse Markers in Speech:
Characteristics and Challenges for Corpus Annotation
Ludivine Crible LUDIVINE.CRIBLE@UCLOUVAIN.BE
Université Catholique de Louvain
Place Blaise Pascal, 1
1348 Louvain-la-Neuve Belgium
Maria-Josep Cuenca MARIA.J.CUENCA@UV.ES
Universitat de València
Facultat de Filologia
Av. Blasco Ibánez 32
46010 València Spain
Editor: Manfred Stede
Submitted 04/2016; Accepted 10/2017; Published online 12/2017
Abstract
It is generally acknowledged that discourse markers are used differently in speech and writing, yet
many general descriptions and most annotation frameworks are written-based, thus partially unfit to
be applied in spoken corpora. This paper identifies the major characteristics of discourse markers in
spoken language, which can be associated with problems related to their scope and structure, their
meaning and their tendency to co-occur. The description is based on authentic examples and is
followed by methodological recommendations on how to deal with these phenomena in more
exhaustive, speech-friendly annotation models.
Keywords: discourse markers, corpus annotation, speech, linguistic complexity, mode of
communication
1 Introduction
Within the large literature on discourse markers (henceforth DMs), a number of studies investigate
how uses differ according to the mode of communication (see for instance Chafe, 1982; Horowitz
and Samuels, 1987; Castellà, 2004; Biber, 2006; López Serena and Borreguero Zuluoga, 2010).
Yet corpus-based analysis of spoken discourse is still quite infrequent, or focused on very specific
phenomena (see López Serena and Borreguero Zuluoga, 2010; Ciabarri, 2013; Cuenca, 2013a;
Crible, 2017 for some exceptions). Similarly, most annotation models of DMs are designed for
written discourse. This is the case of the Rhetorical Structure Theory or RST (Mann and Thompson,
1988), the Penn Discourse Treebank or PDTB 2.0 (Prasad et al., 2008) and the Cognitive approach
to Cognitive Relations or CCR (Sanders et al., 1992) – although there are some recent endeavors
to apply these frameworks to speech (e.g., Tonelli et al., 2010 for PDTB in spoken Italian). Despite
this predominance of writing-based models, spoken and spoken-like corpora are increasingly
available with large databases such as the Spoken BNC2014 project for English or C-Oral-Rom for
Romance languages. Therefore, this paradoxical situation calls for the need to take into account the
DISCOURSE MARKERS IN SPEECH
150
particular characteristics of spoken language both for descriptive adequacy and in order to extend
writing-based annotation models, or even to design multimodal models (Crible and Zufferey,
2015).
This paper aims at identifying the major characteristics of discourse markers that relatively
differ (in nature, frequency or degree) between spoken and written language, and discusses the
problems that they pose for corpus annotation. Besides examples and some corpus-based
quantitative findings of the distribution of these features, the paper also provides a review of how
existing writing-based frameworks handle (or not) the particular characteristics of spoken DMs.
This qualitative discussion of the challenges of spoken discourse annotation will lead to
methodological recommendations to enhance the validity of future corpus-based research.
2 Spoken vs. Written Mode of Communication and the Use of DMs
In this section, we report on previous descriptive works which identified relative differences
between the spoken and written modes of communication (Section 2.1), and their resulting impact
on the use of DMs (Section 2.2).
2.1 Characteristics of Speech and Writing
It seems uncontroversial that written and oral texts are produced and processed differently (see,
e.g., Horowitz and Samuels, 1987). As a consequence, DMs are expected to behave differently
according to the mode of communication. However, linguistic complexity in speech and writing is
a controversial matter.
Two opposite positions can be distinguished. Chafe (1982) and many other authors consider
that oral grammar is simpler than written grammar. Chafe argues that there are two dichotomies
that account for this fact: syntactic fragmentation vs. integration and detachment vs. involvement.
Halliday (1987) represents the opposite, and less popular, perspective. Halliday argues that the
grammar of speech is complex and so the idea that speech is structureless is a myth. “Speech is not,
in any general sense, ‘simpler’ than writing; if anything, it is more complex” (Halliday, 1987: 65).
What seems to be uncontroversial is that spoken and written language differ as for their respective
kind of complexity. Written language tends to be lexically dense but grammatically “simple” or
rather “regular”. Conversely, speech is grammatically intricate but lexically sparse, as illustrated
by the frequent cases of underspecification, implicitness and/or ambiguity in spoken language:
“speech [is] processlike, intricate, with meanings related serially” (Halliday, 1987: 79-80). Their
preferred patterns of organization are therefore different.
An important methodological assumption (or conclusion) of Halliday’s perspective is that the
term mode of communication must be understood in a wide sense: mode is not only (or mainly) the
medium, but the contextual factors attached to it.1 Although the mode of communication, either
spoken or written, is an important factor to account for the selection and frequency of discourse
markers, other factors, namely planning and interactivity, must also be taken into account.
Planning is more likely to be carried out in written texts. Planning is directly related to the
following features: “face-to-text”, not space-time bound, secondary discourse and consciousness.
It can also be directly associated with formality, since informal texts are usually not planned,
whether they are oral or written. That is why an informal and non-revised text by a non-expert often
sounds like spoken language. The other side of planning is revision once the message has been
textualized. Revision is only possible with permanent products.
Interactivity, that is, the presence and activity of the hearer explicitly expressed from the
speaker’s point of view, is more likely to occur in spoken texts or, rather, in face-to-face interaction,
1 The same position is put forward in Bazzanella (2001).
DISCOURSE MARKERS IN SPEECH
151
including “speech-like” modes of mediated communication (e.g. chats). Interactivity is related to
face-to-face communication, addressor-addressee reciprocity and interpersonal discourse, as
opposed to objective and distant communication. It can be related to an additional factor, namely,
emotivity. Emotivity is typically linked to orality, informal communication and unplanned texts
(see Cuenca and Romano, 2013). In fact, Koch and Österreicher (1990: 12) identify intimacy and
emotional implication as factors defining communicative proximity (vs. affective distance and lack
of emotional implication).2
In short, speech is not always unplanned and interactive (e.g. a broadcast political speech) and
writing is not necessarily planned and non-interactive (e.g. text-based chats). The opposition
between speech and writing should therefore be refined by taking these contextual factors into
account, so that we will henceforth always specify what particular type of discourse is at stake.
2.2 DMs in Speech vs. Writing
A number of previous analyses of the linguistic complexity of oral and written texts have taken into
account the use of DMs. In this section, two of them will be reviewed, namely Castellà (2004) and
Biber (2006).
Castellà (2004) analyzes three discourse genres in Catalan (informal conversation, academic
lecture and written academic prose) by focusing on three aspects of linguistic structure and
complexity: lexical density, compound sentences, and text connection. Castellà’s main conclusion
is that spoken informal texts, spoken formal texts (expository) and written formal texts result from
different discourse-building strategies (2004: 149). This fact has an impact on both the linguistic
complexity of the discourse and the use of DMs. Specifically,
Spoken conversation exhibits a verbal style (more verbs and verbal complements) whereas
written exposition has a more nominal style (more nouns and noun complements). As a
consequence, written prose organizes information through integration of structures
(complex phrases, mainly noun phrases, and clauses), while conversation includes more
reduced phrases and clauses, flexible structures and more repetition. Writing is, then,
denser and more lexically varied.
Written texts tend to relate text units as compound sentences, whereas spoken texts make
a profuse use of text connection.
Orality combines intonation, sentential connection and text connection to link units.
Writing, in contrast, heavily relies on sentential connection alone.
All texts, whether written or spoken, use a similar number of subordination constructions.
Informal conversations include a similar number of subordinators in absolute terms and in
relation with the number of words, although fewer if we consider the total number of
clauses. This puzzling fact (the paradox of complexity, as Castellà names it) has to do with
the tendency to an increase of subordinators because of two different preferred strategies:
a verbal style, typical of speech, implies connection between verbal units; similarly, an
integrated style, typical of writing, relies on sentential connection.
As for the use of DMs, some conclusions can be drawn from Castellà’s analysis:
Informal conversation resorts more often to conjunctions and all kinds of DMs as text
connectives. Spoken expository prose is more similar to written prose in the use of
conjunctions as text connectives, but to informal conversation in the use of other
connectives.
Spoken genres use a more reduced variety of connectives, but they are more polyfunctional.
2 In the words of Tannen (1982: 15), interpersonal involvement characterizes orality, whereas written texts are more
message-oriented.
DISCOURSE MARKERS IN SPEECH
152
Some connectives and some uses of certain connectives seem to be typical of one type of
text.
Douglas Biber, in his monograph on University Language (2006), analyzes a wide range of the
registers found in American universities, both spoken (classroom teaching, classroom management,
office hours, study groups, service encounters) and written (academic textbooks, course
management, institutional writing). As for the use of DMs, Biber (2006: chapter 4) concludes that:
Linking adverbials, such as however, therefore, for example, that is, are found in both
spoken and written registers but “are primarily characteristic of written registers” (2006:
70), mainly of textbooks.
“Discourse markers” (such as okay, well, now) are “restricted primarily to spoken
discourse” (2006: 66).
As for dependent clauses, “overall patterns of use […] run exactly opposite to […] prior
expectations, with dependent clauses overall being more common in the spoken university
registers than in the written registers” (Biber, 2006: 72). Specifically, adverbial clauses and
complement clauses, especially that-clauses, are more frequent in speech, whereas relative
clauses occur in both modes of discourse depending on the genre but are more common in
written registers.
To sum up, our intuition as linguists or even as speakers is that, since written and spoken texts are
produced and processed in a different way, discourse markers should exhibit a different behavior
accordingly. However, previous studies have shown that the analysis of written and spoken texts
does not lead to a straightforward conclusion. Orality is not a homogeneous phenomenon and thus
there are many intermediate cases between the pole represented by (prototypical) spoken texts and
(prototypical) written texts. Similarly, it is possible to identify spoken texts with a structure and use
of DMs that resemble that of written texts, and also written texts that echo orality. Nonetheless, the
analysis of spoken texts shows that the use of DMs tends to exhibit some relative differences or
tendencies. This is especially the case in spontaneous speech, where planning is low, and in
dialogue, where interactivity is high.
3 Materials and Present Approach to DMs
In the following sections, some characteristics of spoken DMs will be discussed and illustrated by
naturally-occurring examples taken from DisFrEn, a French-English comparable corpus fully
annotated for discourse markers (Crible, 2017b). This dataset comprises 15 hours of recordings
(161,700 words) balanced across a variety of more or less formal settings such as conversations,
interviews or political speeches. The English transcripts were compiled from existing corpora,
namely the British component of the International Corpus of English (Nelson et al., 2002) and the
Backbone project (Kohn, 2012). A total of 8,743 DM items have been identified and annotated
following the definition and guidelines in Crible (2017a): DMs were identified through a bottom-
up selection (no closed list) provided they met the criteria of procedurality, syntactic optionality,
fixedness (i.e. grammaticalized), discourse-level scope and metadiscursive function (discourse
relation, topic structure, turn exchange, speaker-hearer relationship). This definition includes in
effect expressions as varied as conjunctions (and, but, although), adverbs (actually, well),
prepositional phrases (in fact, by the way) or verb phrases (you know, I mean, if you will), among
others.
The annotation covers syntactic position, co-occurrence and sense disambiguation. Crible and
Degand (forthcoming) report on moderate inter-annotator agreement score (Fleiss’ kappa = 0.563,
relative agreement 70.9%) while intra-annotator agreement, performed on 10% of the whole
corpus, is higher (𝜅 = 0.74, 75.8% of relative agreement). This data will be used to quantify the
DISCOURSE MARKERS IN SPEECH
153
characteristics under discussion in this paper, comparing, when available, results from available
written corpora (mostly the PDTB corpus of Wall Street Journal, Prasad et al., 2008).
4 DMs in Unplanned Speech: The Scope of DMs and Discourse Structure
The standard use of DMs implies two complete and identifiable arguments, which are consecutively
attached to a DM. However, in speech, and especially in unplanned and/or interactive discourse,
this is not always the case. One of the two arguments can be missing, incomplete or implicit
(truncated structures) and the two arguments can be complex and interrupted by different arguments
(far-reaching scope), so that the linear structure corresponding to the construction “Arg1 DM Arg2”
does not always match its semantic structure.
4.1 Truncated Structures
Structures containing DMs in unplanned speech are often truncated (that is, the second argument
is missing or is not complete) or independent (the first argument is implicit), whereas in
written/planned discourse the structures tend to be continuous and complete because the text is
presented as a finished product (instead of an on-going process in speech, cf. Halliday 1987: 65).
This fragmentation of unplanned spoken syntax affects the annotation process by blurring the
function of DMs in truncated structures and generates incomplete argument pairs. Stent (2000)
already noticed this feature of dialogs in her application of RST to conversational data, which led
her to discard abandoned utterances from the scope of the annotation. She further notes that full
representation of discourse structure as proposed by the RST framework cannot be fully achieved
“given the length and complexity of a typical dialog” (2000: 250). Cases of missing Arg2 and
missing Arg1 are respectively illustrated in Examples (1) and (2).
(1) ICE24: I was hoping it’d be a week so I could get off the ward for a bit but
ICE25: oh already?
ICE12: well you know it’s nice to have a change (EN-phon-05)
(2) ICE25: I’ll walk down anyway so I can get one
ICE26: oh right ok
ICE25: so (0.573) um (1.730) so are you still on call at work (EN-phon-05)
In (1), the second argument of “but” is not verbally expressed, leaving ICE25 to infer what is left
implicit. In (2), ICE25 is struggling to come up with a new topic and uses two tokens of “so” to
hold the floor while planning the upcoming utterance: the argument introduced by “so” bears no
relation with previous context and “so” simply expresses a stalling and topic-shifting function. In
these examples, interactivity plays an essential part since, in both cases, the incomplete structure is
either turn-final (1) or turn-initial (2), pointing to the specific role of turn transitions in DM
behavior.
Such non-relational functions of DMs are excluded from writing-based frameworks such as
RST and PDTB where two abstract objects are required for any sense to be assigned. Yet, absence
of left or right context is quite frequent for spoken DMs, as already mentioned in Degand and
Simon-Vandenbergen (2011: 289), who talk about a “scale” of relationality, from “non-relational”
markers (e.g., actually) to “strictly relational” (e.g., because) discourse markers. Some uses have
become conventionalized as specific discourse functions: Estellés and Pons (2014) identified the
“absolute initial position” which can only be filled by a limited number of DMs in Spanish (e.g.,
bueno ‘well’) that do not require previous context and that are typically used to signal the beginning
of a new interaction. Similarly, final position (as in Example 1 above) is generally associated with
intersubjective, elliptical or common-ground functions that invite the hearer to contribute to the
discourse representation (see Haselow, 2012; Buysse, 2014; Degand, 2014), functions which are
completely absent from most taxonomies of discourse relations. To date, most works on spoken
DISCOURSE MARKERS IN SPEECH
154
corpora (e.g. Tonelli et al., 2010; Rehbein et al., 2016; Riccardi et al., 2016) make a distinction
between “connectives” and “discourse markers”, that is, relational vs. non-relational uses, the latter
being discarded although they are sometimes performed by the same expressions (cf. but and so in
the examples above), thus maintaining a writing-based, strictly relational view of DMs.
4.2 Far-reaching Scope
Structures containing DMs in speech often display long-distance relationships, or “far-reaching
scope” (Lenk, 1998: 247), as a result of the dynamic process of building a shared discourse
representation. This feature is mainly displayed in two configurations: a DM relating to a distant
previous context of which it is separated by a digression (location), or a DM taking scope over
multiple utterances which, once combined, constitute its left argument (extent). This flexibility is
not specific to speech: Unger (1996: 409), for instance, acknowledges that, in writing as well,
“discourse connectives can have scope over an utterance or a group of utterances”. Similarly, in
the PDTB 2.0 (Prasad et al., 2008: 3), the extent and location of Arg1 are coded as separate
variables, amounting to 617/18459 (3.34%) cases where Arg1 comprises multiple utterances and
1666/18459 (9%) cases where Arg1 is non-adjacent to Arg2 (the possibility of a non-adjacent Arg1
is restricted to explicit connectives only).
However, some writing-based taxonomies do not include relations which specifically function
with a far-reaching scope, at a higher level of discourse organization. Feng et al. (2014: 1) oppose
in this regard PDTB-style vs. RST-style annotations: “PDTB-style discourse relations encode only
very shallow discourse structures, i.e., the relations are mostly local, e.g. within a single sentence
or between two adjacent sentences”, as opposed to RST which includes “long-distance discourse
dependency” and higher-level relations such as topic-shift or topic-drift. In other words, according
to the PDTB, the same discourse relation can have a local or global scope, but no relation is specific
to the latter: the only potentially far-reaching relation type, viz. “list”, has been removed from the
latest version of their taxonomy (PDTB3, Webber et al., 2016).3 Similarly, in CCR, relations
between topics such as topic-shift or topic-resuming are excluded on the grounds that they are
“orthogonal to another classification in terms of coherence” (Sanders et al., 2016), that is, any
marker or relation expressing a higher-level function (e.g. topic-shift) necessarily encodes a more
local relation such as contrast or consequence, thus forcing a lexical-semantic interpretation of the
DM regardless of its meaning-in-context. Example (2) above seems to contradict this strong
position, where “so” cannot be related to the previous context by any relation other than topic-shift.
It is, however, true that in some cases, the scope of the DM is not clear and seems to combine
local and global meanings, as illustrated by Example (3) where both location and extent of Arg1
are quite flexible:
(3) BB1: could you talk a little bit about the Wirral accent I I know that um (0.200) there’s
obviously quite a um range of accents in that part of the country
BB4: yeah (0.520) uh well I (0.290) consider myself to have a Cheshire accent because
when I was born (0.300) and I lived in (0.110) on the Wirral (0.287) uh (0.333) i-
(0.460) it was a Cheshire accent which is (0.440) the accent I have now though
(0.270) there are overtones of (0.230) the Liverpudlian accent (0.290) however
over the years certainly it has changed (0.270) and now it’s very much (0.110) a
Liverpool accent (0.340) and uh you know which (0.430) I’m not (0.300) I’m not
saying I disapprove of it but I think it’s a lazy speech and you need to (0.440)
3 It should be noted that, in the PDTB, the “NoRel” (i.e. absence of relation) option may mask the presence of possible
long-distance relations, especially when their identification is unclear, in order to enhance consistency in the annotation.
It remains that the PDTB taxonomy only covers typically local relations such as cause or contrast, and not typically far-
reaching relations such as topic-shift.
DISCOURSE MARKERS IN SPEECH
155
actually um (0.530) think about what you’re saying I know my nephew
sometimes’ll to speak to me in the Liverpool accent (0.350) and I’ll say please
speak to me in English (0.160) but it’s things like “yeah” and “you
what” and (0.230) whereas you know mine is “yes” “pardon” or whatever
I’m a bit old-fashioned in that way so I do find the accent (0.440) is a bit harsh
and it’s interesting that actually that accent is spread out into the (0.270) uh (0.390)
the parts of north Wales that are very near to the Wirral… (EN-intf-03)
In Example (3), the DM “so” introducing “I do find the accent is a bit harsh” can be interpreted in
multiple ways depending on which left context it refers to. The most local interpretation connects
it to the immediate co-text (“I’m a bit old-fashioned in that way”), in which case “so” would signal
a consequence relation. Another possible interpretation is to relate it to the “nephew” anecdote, of
which it introduces a conclusion. More convincingly, the evaluative content of the verb (“find”)
and adjective (“harsh”) in the second argument echoes an earlier evaluative utterance: “I’m not
saying I disapprove of it but I think it’s a lazy speech”. Lastly, “so” could also be interpreted as
introducing the answer to the original question by the interviewer “could you talk a little
bit about the Wirral accent”, in which case its function is to come back to the main topic after a
lengthy digression. While a conservation decision would only annotate the local interpretation
(consequence relation), a more accurate and comprehensive analysis could combine two senses
(consequence and topic-resuming), an option which is available in many annotation models for
both speech and writing.
Complex examples such as (3) tend to show that, while systematic annotation of DM scope
would be particularly challenging (especially in spoken data), speech-friendly taxonomies should
not deny the ability of DMs to express higher-level functions such as coming back to a previous
topic. In the DisFrEn corpus, topic-relating DMs take up 6.07% of all annotated tokens, which
would not be accounted for by writing-based models with a more restricted view of DM functions.
Given that these functions are often expressed by frequent connectives such as and or so, which
are also used in writing, this observation of far-reaching DMs in speech calls for more inclusive
taxonomies for written corpora as well, in order to reach a better comparability across annotated
corpora.
It thus appears that spoken DMs not only tend to take scope over large and distant segments
(as in writing), but can also combine local and global scope simultaneously, making the annotation
process quite complex and calling for specific function types at a higher level of discourse structure
than what is included in local views of coherence (PDTB, CCR).
5 The Meaning of DMs
One of the most commented features of DMs is their multifunctionality (see for instance Mosegaard
Hansen, 2006, 2008; Hummel, 2012). This feature, which can be directly associated with semantic
and pragmatic ambiguity, is especially outstanding in speech, where a limited number of markers
tend to be used with a relatively wide range of meanings and where context is a key-element in
discourse production and processing.
5.1 Vagueness
The meaning of a DM in spoken/unplanned discourse is often (semantically and pragmatically)
vague or not clearly definable, as opposed to the main tendency in written/planned discourse.4 DMs
can be ambiguous in writing as well, but these cases are much more typical of speech, especially
4 Tuggy (1993) defines “vagueness” in cognitive-grammar terms as “meanings which are not well-entrenched but which
have a relatively well-entrenched, elaboratively close schema subsuming them” (1993: 8).
DISCOURSE MARKERS IN SPEECH
156
in face-to-face settings where the speaker might expect the hearer to make use of various contextual
cues (e.g., prosody, gestures) to reconstruct the speaker’s intended meaning, an option which is not
available for readers (Schober and Brennan, 2003). In addition, speech-specific DMs such as well
or French quoi ‘you know’ do not encode a strong core meaning, as opposed to more explicit
expressions in writing such as by contrast or in addition, which are virtually absent in speech: for
instance, in DisFrEn, only one occurrence of in addition was found out of the total 8,743 annotated
DMs, against 165 in the PDTB corpus out of 18,459 tokens.
While not all written DM occurrences are clear and unproblematic, similarly, not all spoken
DMs are vague, and some forms tend to specialize in a given function, such as the very frequent
you know which is mostly used to check for the hearer’s attention and understanding. Yet, two
degrees of difficulty related to vagueness are illustrated in (4) and (5) where the DM I mean departs
from its basic reformulative meaning.
(4) ICE76: didn’t John use to deal with uhm (1.850) divorce in his earlier days
ICE77: did he uhm
ICE76: found it quite depressing
ICE77: I imagine it probably is uhm
ICE76: all the toing and froing and
ICE77: I mean I sh- I should think you’d get over the uhm (0.110) the voyeuristic aspects
in the early stages (EN-conv-04)
(5) ICE12: this is what we do all the time we sit and describe other people and I mean when
people got stuck I’d just say look just list you imagine... (EN-conv-03)
In (4), I mean elaborates on the speaker’s previous turn but it is not clear whether it introduces
more detail (“it probably is depressing, by that I mean that you would get over the voyeuristic
aspects”) or a justification (“it probably is depressing, I say that because you would get over the
voyeuristic aspects”), or a combination of both (see next section on multifunctionality). In (5), I
mean does not fluctuate between two meanings, like in (4), but rather seems completely
underspecified and functions as a semantically void “punctuant” (Vincent, 1993). In addition, its
co-occurrence with and blurs its own functional contribution into the additive meaning of the
conjunction (see section 5 on co-occurrence).
The treatment of cases like (4) and (5) does not only depend on the quality and coverage of the
functional taxonomy used by the annotator. These problems are mostly due to the intrinsic
ambiguity of language in general and of some DMs in particular. The challenge for the annotator
is then to assign the appropriate function label within the taxonomy, striving to avoid “undefined”
or “non-interpretable” categories which usually end up left out of quantitative analyses as outlier
data. Therefore, a marker-based approach to DMs, i.e. starting from the identification of (a set of)
DMs and systematically annotating their function(s), presents the additional challenge to deal with
such vague cases and not only the more typical relational uses (such as I mean expressing
reformulation or specification).
5.2 Polysemy and Multifunctionality of DMs
As already indicated, another challenging characteristic of DMs – in general and especially in
speech – is their multifunctionality, which is at the very core of the category to a greater extent than
more homogeneous pragmatic categories such as modal particles (associated with epistemic
modality), interjections (associated with subjectivity) or response signals (associated with
agreeing/disagreeing). As suggested in Crible (2017a: 107), three forms of multifunctionality can
be distinguished:
(1) the category covers items that perform many different functions; (2) a single member can
perform different functions in different contexts; and (3) a single member can perform different
functions simultaneously in the same context, given the great polysemy of DMs.
DISCOURSE MARKERS IN SPEECH
157
We will now turn to each of these degrees of multifunctionality to discuss their challenge for corpus
annotation.
5.2.1 One Category with Different Functions
DMs can exhibit propositional functions indicating a logico-semantic relationship (e.g.,
consequence), but also non-propositional, mainly structural functions (e.g., topic-shift). In addition,
some DMs add a modal, (inter)subjective value (e.g., agreement, monitoring, emphasis, etc.) to a
structural function. These three categories (viz. propositional, structural, modal), borrowed from
Cuenca (2013a), have a different weight in speech and in writing.
DMs acting at a propositional level are prevalent in planned monologues (36.5% of all DMs in
news broadcasts vs. only 20% in conversations in DisFrEn), which can be taken as evidence for
their higher attraction to written-to-be-spoken and, by extension, written discourse. Regarding
function types, however, there is a common core of propositional relations in speech and writing,
with examples of cause or contrast, as in (6), which are equally included in writing-based and
speech-based taxonomies.
(6) I wasn’t looking forward to do it but I am now it’s going to be good (EN-phon-01)
DMs with structural functions are typical of both speech and writing, but each mode of
communication tends to prefer certain types of structural functions over others. In (spoken and
written) situations with intermediate or high planning, markers of global structure such as topic-
shift or enumeration (Example 7) are quite frequent. In the DisFrEn corpus, news broadcast (highly
prepared, written-to-be-spoken) is the only register where the topic-shift function ranks among the
most frequent functions (5th and 4th in English and French, respectively). On the other hand,
situations with low planning will make more use of local structuring devices such as continuity
(Example 8).
(7) I will begin with a review of the economic situation and prospects (0.380) I shall then deal
with monetary policy and public finances (0.220) finally I will present my tax proposals
(EN-poli-02)
(8) you sit there in a canteen wherever you are and you sit and you listen and the conversation
next to you someone’s saying (EN-conv-03)
Global structural functions, such as discourse beginning, discourse closing or pre-closing, topic
shift or turn change, are generally not taken into account – or not systematically – in writing-based
taxonomies for DM functions, although they are very frequent in speech (14% of all DMs in
DisFrEn; cf. Examples 2, 3, 7). As already mentioned above, only RST includes a category of topic
relations, while the only structural relation in the PDTB is the local conjunction relation (typically
expressed by and as in Example 8).
DMs that bracket units of talk while also performing modal functions are frequent in genres
where interactivity is more prominent such as dialogues. Modal markers introducing units of talk
are typically oral (e.g., well, look, listen). Some of these uses are covered by the question-answer
relation in RST, although not all turn-initial DMs actually belong to such a well-defined structure,
as in (9) where “look” follows an imperative utterance and not an interrogative one, thus expressing
disagreement (modal) as well as turn-opening (structural).
(9) ICE5 have some banana bread
ICE6 look I’m not that much of a banana bread eater (EN-phon-08)
Regardless of the classifying system, authors agree on the multifunctionality of the DM category,
although not always to the same degree of inclusiveness and granularity. As mentioned above,
structural functions are usually excluded from writing-based taxonomies (e.g. PDTB, CCR) on the
grounds that these relations function at a different, higher level of discourse coherence. Modal
functions are, in turn, absent from these frameworks since they fail to meet the relationality criterion
(e.g. a you know does not relate two arguments but merely takes scope over one utterance).
DISCOURSE MARKERS IN SPEECH
158
However, such an exclusion overlooks emerging modal uses of typical DMs such as French mais
‘but’ which, besides its relational functions (contrast, concession) can also be used structurally
and/or modally in speech, as in (10).
(10) c’est fatigant tu vois on prend leur bics pour le TU ben euh (0.900) quoi mais on n’a
pas dit qu’on voulait bien gnagna tu vois
it’s exhausting you know we take their pens for the TU well uh (0.900) what mais ‘but’
we did not say we were okay blabla you know (FR-conv-05)
In this extract, the speaker is pseudo-reporting someone else’s words (“quoi mais on n’a pas dit
qu’on voulait bien”), where “mais” opens the reported segment and expresses disagreement. Once
more, in a DM-based approach, such uses of “mais” cannot be discarded and have to be accounted
for beyond what a purely relational (contrastive) reading would suggest. Furthermore, the inclusion
of structural and modal functions in DM taxonomies is in line with seminal, empirically-evidenced
theories of discourse domains such as Halliday’s (1970) ideational-textual-interpersonal
distinction, taken up by Schiffrin (1987), Redeker (1991) and many others since then.
5.2.2 One Marker with Different Functions in Different Contexts
DMs in speech/unplanned discourse tend to be multifunctional (Mosegaard Hansen, 2008;
Hummel, 2012). While multifunctional or ambiguous expressions do occur in writing/planned
discourse as well (e.g. and), writing tends to rely on a larger diversity of connectives, some of
which have highly specific meanings (e.g. in addition), as opposed to speech where a few very
generic markers are used extensively for many different functions. In DisFrEn, spoken registers
with a high degree of preparation (political speeches, news broadcasts) show the highest type-token
ratio of DMs (around 20 vs. only 6 in unplanned speech such as casual conversations). These
corpus-based findings suggest that DM expressions are more varied in writing and in planned
“written-to-be-spoken” discourse (see also Cuenca 2013b).
A case in point is the DM and. The DM and combines the problematic features of
multifunctionality and vagueness or underspecification, a situation which, coupled with its high
frequency in speech (the most frequent English DM in DisFrEn), makes it very complex to annotate
reliably. And is quite pervasive in writing as well though not to the same extent as in speech: and
takes up 26% of all DM occurrences in DisFrEn, as opposed to only 16% in the PDTB corpus.
Table 1 compares the most frequent functions of and (more than 10 occurrences) in DisFrEn and
in the PDTB corpus and their proportions over the total number of and tokens. We see that the
functional range of and in the spoken corpus is much larger and more balanced than in writing,
where only four different senses show more than 10 occurrences with an overwhelming majority
of the basic conjunction relation.
Although the two corpora have been annotated with different taxonomies in potentially
different degrees of granularity, both spoken and written and share a number of common functions:
addition roughly corresponds to conjunction, enumeration to list and consequence to result. Other
functions expressed by and in speech are included in the PDTB taxonomy but were not assigned to
and in the written corpus, namely specification, temporal and contrast. Nevertheless, it is obvious
that not all functions in the spoken corpus can be reduced to these core relational meanings: at least
the structural functions of topic-shift and opening boundary must be differentiated from the basic
meaning of and as a conjunction at sentence level (and its underspecified, contextually enriched
uses such as contrast or temporality). While this latter group of structural functions is not part of
the PDTB taxonomy, it remains that the majority of the labels attributed to and in speech (6 out of
11) are included in both annotation models, and yet are not distributed in the same way across the
two modes of communication: more balanced proportions across more different function types in
speech, overwhelming monopoly of one sense label (namely conjunction) in writing.
DISCOURSE MARKERS IN SPEECH
159
DisFrEn (speech): 1140 and PDTB (writing): 3000 and
Addition (57.11%) Conjunction (90.76%)
Specification (15.79%) List (7%)
Consequence (8.86%) Result (1.26%)
Topic-shift (3.6%) Juxtaposition (0.36%)
Temporal (2.37%)
Punctuation (2.11%)
Conclusion (1.75%)
Topic-resuming (1.4%)
Contrast (1.14%)
Opening boundary (1.14%)
Enumeration (0.88%)
Table 1. Most frequent functions of and in speech vs. writing
In other words, these corpus-based results show that DMs tend to fulfil a wider functional range
in speech than in writing, even for those expressions which do occur in both modalities and which
share a common relational core such as and. Besides the inclusion of more different types of
functions, this difference also suggests a greater complexity of spoken DM annotation where the
annotators have to choose among many different possible senses for a single DM, including highly
underspecified uses.
5.2.3 One Marker with Simultaneous Functions in the Same Context
The third level of multifunctionality refers to the tendency of DMs to perform more than one
function at a time (see Petukhova and Bunt, 2009). It is generally agreed (e.g., Brinton, 1996: 35)
that it is not always possible nor relevant to rank by order of semantic priming the respective weight
of multiple functions in a single occurrence. These simultaneous functions can either belong to the
same level or domain (11) or to different ones (12-13).
(11) these concerns are especially important (0.270) as we approach the crucial topic of
economic and monetary union (EN-poli-01)
(12) and then there are the really bland ones that I think oh come on (0.130) you know (0.493)
friendly (0.960) helpful (EN-conv-03)
(13) and that means either we’re going to have to increase the amount of tax we pay (0.400)
and none of us really wants to do that either (0.560) or we’re going to have to reduce
expenditure (EN-intf-04)
In example (11) “as” exhibits two functions from the propositional domain, namely a causal and a
temporal meaning. Similarly, “you know” in Example (11), expresses both monitoring (modal) and
specification (propositional), whereas “either…or” in (12) express alternative (propositional) and
enumeration (structural).
Simultaneous functions of DMs are not more frequent in unplanned than in planned discourse:
only 4% of all DMs in DisFrEn were assigned double labels, an average which is not particularly
sensitive to variation in degrees of planning or interactivity (4.57% in conversations vs. 4.56% in
political speeches). By contrast, results from written corpora tend to show a high frequency of
double senses: for instance, Webber et al. (2016) report that 1002 out of 4138 relations between
conjoined verb phrases are assigned two or three senses in the PDTB corpus.
Although this characteristic exists both in speech and writing, it involves different types of
functions: in the PDTB, for instance, double senses only correspond to simultaneous discourse
relations (e.g. conjunction-synchrony; see also Baldridge and Lascarides, 2005), whereas the
multidimensionality of spoken communication is particularly prone to expressions working on
DISCOURSE MARKERS IN SPEECH
160
different levels or macro-functions at the same time. As Bunt (2012) puts it:
The phenomenon of multifunctionality can be explained by considering participation in a
dialogue as involving multiple activites at the same time, such as making progress in a
given task or activity; monitoring attention and understanding; taking turns; managing
time, and so on. (Bunt 2012: 243)
Once again, we see that writing-based taxonomies of discourse relations, in particular those
excluding topic relations and modal meanings, would fail to account for the full multifunctionality
of spoken DMs “participating in a network of ideational, interpersonal, and textual relations”
(Goutsos 1996: 167).
6 Co-occurrence of DMs
DMs have been repeatedly shown, especially in speech-based research (e.g. Bazzanella, 2001;
Pons, 2008; Cuenca and Marín, 2009; Dostie, 2013; but see also Fraser 2013 on writing), to
frequently co-occur with each other. DM co-occurrence is very frequent in spoken language: up to
20% of all DMs in DisFrEn were coded as part of a co-occurring string of DMs, in combinations
as varied as and so, but I mean, because if or and actually. To our knowledge, such information is
not made available in discourse-annotated corpora. The PDTB reports on a different yet related
phenomenon, which is the coordination of conjoined connectives: only before and after, if and
when and when and if are attested for a total of five occurrences expressing combinations of
temporal and conditional relations. More varied forms and functions of co-occurrence may well
exist in writing, yet further empirical evidence is needed to effectively compare the proportion of
co-occurring DMs across spoken and written corpora.
While comparisons of frequency are limited, differences remain as to the treatment of co-
occurring DMs in speech- vs. writing-based annotation frameworks: in the latter, co-occurring DMs
are treated as one complex unit or distinct type with only one sense label assigned to them; in the
former, however, authors tend to acknowledge varying degrees of fixation, from simply juxtaposed
DMs (in which case each component expresses a distinct function) to more and more bound
combinations expressing one single function. With time and spread use, clustered DMs can indeed
aggregate and combine into new “complex” DMs which are no longer separable and express a
meaning that is not computable from their individual components (Waltereit, 2007), such as and
then and its French equivalent et puis. Yet this is not systematically the case and this variation
should be accounted for more systematically during DM identification and sense disambiguation.
One such proposal is provided by Cuenca and Marín’s (2009) three-fold distinction between
juxtaposition of DMs (different functions generally with different scopes), addition of DMs
(different functions with the same scope) and composition of DMs (a single function performed by
a complex unit whose components are not completely lexicalized).
Juxtaposition of markers implies that two or more markers co-occur but do not combine neither
syntactically nor semantically. In many cases of co-occurrence, the two (or more) DMs remain
“simply” juxtaposed, as in Example (14), where the meanings and functions of three DMs are
serially added to each other.
(14) so it’s actually a proper increasing function (2.833) ok (1.730) so for example if you
wanted to supposing you’re looking at sine x (EN-clas-04)
In (14), three different cases of connection can be identified, namely, consequence, exemplification
and condition.
On the other hand, some co-occurrences imply some degree of integration of the components.
Addition of markers implies two or more markers which combine and have the same scope. They
usually act at a local level but still maintain their meanings and functions sufficiently distinct, as in
(15).
DISCOURSE MARKERS IN SPEECH
161
(15) he’s the guy who is supposed to have left and he had my papers and so that was the
problem over the party (EN-conv-06)
In this example, “and” indicates its basic meaning of addition, marking the continuity of the
following segment, while consecutive “so” specifies that this next segment (“that was the
problem”) is a conclusion to the previous context.5
Finally, composition of markers takes place when two co-occurring markers jointly contribute
to indicating a single discourse function at a global level, as in (16).
(16) and she’s feeding a baby (0.200) so uhm (0.200) and then yes of course this is a some
sort of love scene going on (EN-clas-05)
In (16), “and then” acts as a complex marker indicating the continuation of the description of a
situation. This case can be differentiated from other cases in which and and then can be attributed
distinct functions, generally with then acting as a temporal marker, as in (17).
(17) I really found myself enjoying the first ten minutes and then suddenly (0.273) everything
seemed to disintegrate (EN-intr-01)
The most integrated co-occurrences include DMs that indicate interactional values while delimiting
main units of the narrative: the beginning, the end or a major transition place (turn, sequence and
subsequence, enunciation). Other examples of this type are English well then when it is used as
opening-up closings or okay then indicating a discourse boundary and marking continuation,
although there were no occurrences in DisFrEn. Complex DMs are usually language-specific and
even variety-specific (e.g., ou sinon in Belgian French, see Crible, 2015), although some are present
cross-linguistically, such as and then and its French equivalent et puis (also existing in writing).
The scale from juxtaposition to composition is in constant evolution, which is why lexicalization
criteria are needed to draw the line between different stages of integration, following e.g.,
Himmelmann (2004).
In sum, DM co-occurrence is a pervasive and multi-faceted phenomenon in spoken language
and should therefore be included in annotation frameworks in a more systematic yet flexible way
than what is currently available in writing-based models, where it is either ignored or highly
restricted. This difference of treatment in the literature might not necessarily reflect a relative
difference of frequency in corpus data. Nevertheless, DM co-occurrence appears to be one area
where studies on spoken language can enrich annotation models designed for both speech and
writing.
7 Conclusion and Recommendations for Corpus Annotation
Some degree of DMs specialization depending on the mode of communication (spoken vs. written)
can be observed, but planning and interactivity also play a crucial role on the use of DMs. Some
DM configurations are typically oral and are usually related to unplanned and interactive discourse,
while others are typical of written texts or, rather, planned discourse, either in written or spoken
mode. Overall, it appears that markers in speech are associated with more varied structural
configurations and multifunctionality at various levels (or “multidimensionality” in Petukhova and
Bunt’s (2011) terms), and thus potentially more complex to annotate. As a consequence, annotation
instructions should be adapted/formulated according to the characteristics of spoken DMs in a
systematic and comprehensive way, taking into account a number of relative tendencies which we
have identified, quantified and illustrated.
The structure of speech is linearly intricate, lexically vaguer and includes more repetition than
written discourse. Interactivity and low planning in dialogal genres are prone to repair, turn-taking
5 The recurrence of and so in DisFrEn (26 occurrences) can be a telling sign of its increasing fixation as a new combined
unit in language.
DISCOURSE MARKERS IN SPEECH
162
and overlapping. This implies the presence of truncated or independent clauses as well as long-
distance relationships which pose a challenge to the interpretation of meaning and thus to
annotation. In addition, DMs often seem to perform vague functions. Structural and modal
functions (and combinations thereof) must be correctly identified and included in annotation
taxonomies.
As a conclusion, we suggest the following recommendations for spoken corpus annotation to
handle the problematic cases identified:
a) The absence of a textually expressed Arg1 or Arg2 must be considered as a structural
possibility. It covers both non-relational DMs (that is, applying to one segment only) or
relational DMs applying to at least one implicit segment such as a pragmatically available
assumption or extra-linguistic context.
b) Longer excerpts must be taken into consideration to identify the scope of the DM.
Simplified paraphrases can be used to identify the core structure. However, this is
challenging to annotate consistently, and explicitly identifying the units under a DM’s
scope may be too ambitious, especially in long-distance relations (for both spoken and
written data). We argue that sense disambiguation is informative and complex enough and
should not necessarily be combined with an identification of the related segments.
c) Most DMs can act on different levels or domains, that is, they can implement propositional
and non-propositional meanings. The “one form-one function” principle seldom applies.
Structural and modal functions should be incorporated in general DM taxonomies. Specific
disambiguating criteria should be established for operational annotation.
d) Double labels should be allowed. However, they must be restricted to true simultaneous
multifunctionality and not used in case of hesitation between two senses. Whether or not
double senses (e.g. causal-temporal) constitute new “complex” types of functions should
be left to the analyst’s choice depending on the method and research question.
e) Given the complexity of discourse annotation in spoken data, a tracking system to
document the annotator’s confidence during the process would be useful in order to retrieve
ambiguous cases, and potentially discard them or at least distinguish them in the analysis.
f) DM co-occurrences should be differentiated and systematically annotated. Various degrees
of integration must be identified and taken into account.
In conclusion, by identifying characteristics of DMs which are more frequent (if not only existing)
in speech and by suggesting annotation recommendations, the research presented in this paper aims
to make a contribution to a better description of the use of DMs in speech, and also to the
development of annotation systems that can deal not only with written but also with spoken data.
This would allow us to build inter-operable databases compiling evidence from multiple data types,
thus considerably furthering our knowledge on the use and functions of this complex pragmatic
category.6
Acknowledgements
The first author benefits from the support of the ARC-project “A Multi-Modal Approach to Fluency
and Disfluency Markers” granted by the Fédération Wallonie-Bruxelles (12/17-044). The second
author is part of the Grampint project (FFI2014-56258-P) supported by the Spanish Ministry of
6 This research is in line with the IS1312 COST Action “TextLink: Structuring Discourse in Multilingual Europe”
(http://textlink.ii.metu.edu.tr/, chair Prof. Liesbeth Degand), whose main aim is to map different annotation models into
an integrated interface which bridges differences between languages and theoretical frameworks. Both authors are
members of this COST Action.
http://textlink.ii.metu.edu.tr/
DISCOURSE MARKERS IN SPEECH
163
Education and Science. We thank the three anonymous reviewers and the editors for their insightful
comments.
References
Jason Baldridge and Alex Lascarides (2005). Annotating discourse structure for robust semantic
interpretation. In Proceedings of the 6th International Workshop on Computational Semantics,
pages 17-29, The University of Tilburg, Tilburg.
Carla Bazzanella (2001). Segnali discorsivi nel parlato e nello scritto. In M. Dardano, A. Pelo, A.
Stefinlongo (eds.), Scritto e Parlato. Metodi, Testi e Contesti, pages 79-97, Aracne, Roma.
Douglas Biber (2006). University Language: A Corpus-Based Study of Spoken and Written
Registers. John Benjamins, Philadelphia, PA.
Laurel Brinton (1996). Pragmatic Markers in English. Grammaticalization and Discourse
Functions. Mouton de Gruyter, New York, NY.
Harry Bunt (2012). Multifunctionality in dialogue. Computer Speech and Language 25:222-245.
Lieven Buysse (2014). “We went to the restroom or something”. General extenders and stuff in the
speech of Dutch learners of English. In J. Romero-Trillo (ed.), Yearbook of Corpus Linguistics
and Pragmatics: New Empirical and Theoretical Paradigms, pages 213-237, Springer, Berlin.
Josep M. Castellà (2004). Oralitat i Escriptura. Dues Cares de la Complexitat del Llenguatge.
Publicacions de l’Abadia de Montserrat, Barcelona.
William Chafe (1982). Integration and involvement in speaking, writing and oral literature. In D.
Tannen and R. Freedle (eds.), Spoken and Written Language, pages 83-113, Academic Press,
New York, NY.
Federica Ciabarri (2013). Italian reformulation markers: a study on spoken and written language.
In C. Bolly and L. Degand (eds), Across the Line of Speech and Writing Variation, pages 113-
127, Presses universitaires de Louvain, Louvain-la-Neuve.
Ludivine Crible (2015). Grammaticalisation du marqueur discursif complexe ou sinon dans le
corpus de SMS belge: spécificités sémantiques, graphiques et diatopiques. Le Discours et la
Langue 7(1):181-200.
Ludivine Crible (2017a). Towards an operational category of discourse markers: A definition and
its model. In C. Fedriani and A. Sanso (eds), Discourse Markers, Pragmatics Markers and
Modal Particles: New Perspectives, pages 101-126, John Benjamins, Amsterdam.
Ludivine Crible (2017b). Discourse markers and (dis)fluencies in English and French: Variation
and combination in the DisFrEn corpus. International Journal of Corpus Linguistics 22(2): 242-
269.
Ludivine Crible and Liesbeth Degand. Reliability vs. granularity in discourse annotation: What is
the trade-off? Corpus Linguistics and Linguistic Theory, forthcoming.
Ludivine Crible and Sandrine Zufferey (2015). Using a unified taxonomy to annotate discourse
markers in speech and writing. In H. Bunt (ed.), Proceedings of the 11th Joint ACL-ISO
Workshop on Interoperable Semantic Annotation (isa-11), April 14th, London, UK, pages 14-
22.
Maria-Josep Cuenca (2013a). The fuzzy boundaries between modal and discourse marking. In L.
Degand, B. Cornillie and P. Pietrandrea (eds), Discourse Markers and Modal Particles.
Description and Categorization, pages 191-216, John Benjamins, Amsterdam.
Maria-Josep Cuenca (2013b). Causal constructions in speech. In C. Bolly and L. Degand (eds.),
Text-Structuring. Across the Line of Speech and Writing Variation, pages 17-31, Presses
Universitaires de Louvain, Louvain-la-Neuve.
Maria-Josep Cuenca and Maria-Josep Marín (2009). Co-occurrence of discourse markers in
Catalan and Spanish oral narrative. Journal of Pragmatics 41:899-914.
DISCOURSE MARKERS IN SPEECH
164
Maria-Josep Cuenca and Manuela Romano (2013). Discourse markers, structure and emotionality
in oral narratives. Narrative Inquiry 23:344-370.
Liesbeth Degand (2014). “So very fast, very fast then” Discourse markers at left and right periphery
in spoken French. In K. Beeching and U. Detges (eds), The Role of the Left and Right Periphery
in Semantic Change: Crosslinguistic Investigations of Language and Language Change, pages
151-178, Brill, Leiden.
Liesbeth Degand and Anne-Marie Simon-Vandenbergen (2011). Grammaticalization and (inter-)
subjectification of discourse markers. Linguistics 49:287-294.
Gaëtane Dostie (2013). Les associations de marqueurs discursifs - De la cooccurrence libre à la
collocation. Linguistik Online 62(5).
Maria Estellés and Salvador Pons (2014). Absolute initial position. In S. Pons (ed.), Discourse
Segmentation in Romance Languages, pages 121-155, John Benjamins, Amsterdam.
Bruce Fraser (2013). Combinations of contrastive discourse markers in English. International
Review of Pragmatics 5:318-340.
Vanessa Wei Feng, Ziheng Lin and Graeme Hirst (2014). The Impact of Deep Hierarchical
Discourse Structures in the Evaluation of Text Coherence. In Proceedings of the 25th
International Conference on Computational Linguistics (COLING 2014), pages 940-949,
Dublin City University and Association for Computational Linguistics, Dublin.
Dionysos Goutsos (1996). Modeling Discourse Topic. Sequential Relations and Strategies in
Expository Text. Ablex, Norwoord, NJ.
Michael A. K. Halliday (1970). Functional diversity in language as seen from a consideration of
modality and mood in Englih. Foundations of Language: International Journal of Language
and Philosophy 6:322-361.
Michael A. K. Halliday (1987). Spoken and written modes of meaning. In R. Horowitz and S. J.
Samuels (eds.), Comprehending Oral and Written Language, pages 55-82, Academic Press,
New York, NY.
Alexander Haselow (2012). Subjectivity, intersubjectivity and the negotiation of common ground
in spoken discourse: Final particles in English. Language and Communication 32:182-204.
Nikolaus Himmelmann (2004). Lexicalization and grammaticization: Opposite or orthogonal? In
W. Bisang, N. Himmelmann and B. Wiemer (eds), What Makes Grammaticalization? A look
from its Fringes and its Components, pages 21-42, Mouton de Gruyter, Berlin.
Rosalind Horowitz and S. Jay Samuels (1987). Comprehending oral and written language: Critical
contrasts for literacity and schooling. In R. Horowitz and S. J. Samuels (eds.), Comprehending
Oral and Written Language, pages 1-52, Academic Press, San Diego, CA.
Martin Hummel (2012). Polifuncionalidad, Polisemia y Estrategia Retórica. Los Signos
Discursivos con Base Atributiva entre Oralidad y Escritura. Acerca de esp. bueno, claro, total,
realmente, etc. De Gruyter, Berlin.
Peter Koch and Wulf Österreicher (1990). Gesprochene Sprache in der Romania: Französisch,
Italienisch, Spanisch. Max Niemeyer, Tübingen.
Kurt Kohn (2012). Pedagogic corpora for content and language integrated learning. Insights from
the BACKBONE project. The Eurocall Review 20(2).
Uta Lenk (1998). Discourse markers and global coherence in conversation. Journal of Pragmatics
30:245-257.
Araceli López Serena and Margarita Borreguero Zuluoga (2010). Los marcadores del discurso y la
variación lengua hablada vs. lengua escrita. In Ó. Loureda Lamas and E. Acín Villa (eds.), Los
Estudios de Marcadores del Discurso en Español, Hoy, pages 415-496, Arco/Libros, Madrid.
William Mann and Sandra Thompson (1988). Rhetorical Structure Theory: Toward a functional
theory of text organization. Text 8(3):243-281.
DISCOURSE MARKERS IN SPEECH
165
Maj-Britt Mosegaard Hansen (2006). A dynamic polysemy approach to the lexical semantics of
discourse markers (with an exemplary analysis of French toujours). In K. Fischer (ed.),
Approaches to Discourse Particles, pages 21-41, Elsevier, Amsterdam.
Maj-Britt Mosegaard Hansen (2008). Particles at the Semantics/Pragmatics Interface: Synchronic
and Diachronic Issues. A Study with Special Reference to the French Phrasal Adverbs. Elsevier,
Oxford.
Gerald Nelson, Sean Wallis and Bas Aarts (2002). Exploring Natural Language: Working with the
British Component of the International Corpus of English. John Benjamins, Amsterdam.
Volha Petukhova and Harry Bunt (2009). Towards a multidimensional semantics for discourse
markers. In Proceedings of the 8th International Conference on Computational Semantics
(IWCS-8), pages 157-168, University of Tilburg, Tilburg.
Salvador Pons (2008). La combinación de marcadores del discurso en la conversación coloquial:
interacciones entre posición y función. Estudos Linguísticos/Linguistic Studies 2:141-159.
Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Miltsakaki, Livio Robaldo, Aravind Joshi and
Bonnie Webber (2008). The Penn Discourse TreeBank 2.0. In Proceedings of LREC, June 2008,
Marrackech, Morroco, pages 2961-2968.
Gisela Redeker (1991). Linguistic markers of discourse structure. Linguistics 29:1139-1172.
Rehbein, Ines, Merel Scholman and Vera Demberg (2016). Annotating discourse relations in
spoken language: A comparison of the PDTB and CCR frameworks. In Proceedings of LREC,
May 2016, Portoroz, Slovenia, May, pages 23-28.
Riccardi, Giuseppe, Evgeny A. Stepanov and Shammur Absar Chowdhury (2016). Discourse
connective detection in spoken conversations. In Proceedings of 2016 IEEE International
Conference on Acoustics, Speech and Signal Processing (ICASSP), March 2016, Shangai,
China, pages 6095-6099.
Ted Sanders, Wilbert Spooren and Leo Noordman (1992). Toward a taxonomy of coherence
relations. Discourse Processes 15:1-35.
Ted Sanders, Vera Demberg, Jacqueline Evers-Vermeul, Jet Hoek, Merel Scholman, Sandrine
Zufferey (2016). Unifying dimensions in discourse relations: How various annotation
frameworks are related. In Proceedings of the Second TextLink Action Conference, April 2016,
Budapest, Hungary, pages 110-112.
Deborah Schiffrin (1987). Discourse markers. Cambridge University Press, Cambridge.
Michael Schober and Susan Brennan (2003). Processes of interactive spoken discourse: The role
of the speaker. In A. Graesser, M. Gernsbacher and S. Goldman (eds), Handbook of Discourse
Processes, pages 123-164, Lawrence Erlabum, Hillsdale, NJ.
Amanda Stent (2000). Rhetorical structure in dialog. In Proceedings of the 2nd International
Natural Language Conference (INLG’2000), pages 247-252.
Deborah Tannen (1982). The oral/literate continuum in discourse. In D. Tannen (ed.), Spoken and
Written Language, pages 1-16, Ablex, Norwood, NJ.
Sara Tonelli, Giuseppe Riccardi, Rashmi Prasad and Aravind Joshi (2010). Annotation of discourse
relations for conversational spoken dialogs. In Proceedings of the Seventh International
Conference on Language Resources and Evaluation (LREC 10), pages 2084-2090.
David Tuggy (1993). Ambiguity, polysemy, and vagueness. Cognitive Linguistics 4(3):273-290.
Christoph Unger (1996). The scope of discourse connectives: implications for discourse
organization. Journal of Linguistics 32:403-438.
Diane Vincent (1993). Les Ponctuants de la Langue et Autres Mots du Discours. Nuit Blanche
Éditeur, Québec.
Richard Waltereit (2007). À propos de la genèse diachronique des combinaisons de marqueurs.
L’exemple de bon ben et enfin bref. Langue Française 154:94-128.
DISCOURSE MARKERS IN SPEECH
166
Bonnie Webber, Rashmi Prasad, Alan Lee and Aravind Joshi (2016). Discourse Annotation of
Conjoined VPs. In Conference Handbook of the Second Action Conference of TextLink, April
2016, Budapest, Hungary, pages 135-140.