<!DOCTYPE article PUBLIC "-//NLM//DTD JATS (Z39.96) Journal Archiving and Interchange DTD v1.0 20120330//EN" "JATS-archivearticle1.dtd">
<article xmlns:xlink="http://www.w3.org/1999/xlink">
  <front>
    <journal-meta />
    <article-meta>
      <title-group>
        <article-title>LICO: A Lexicon of Italian Connectives</article-title>
      </title-group>
      <contrib-group>
        <aff id="aff0">
          <label>0</label>
          <institution>Fondazione Bruno Kessler University of Pavia Fondazione Bruno Kessler University of Potsdam University of Pavia</institution>
          ,
          <addr-line>Italy Pavia</addr-line>
          ,
          <country country="IT">Italy</country>
          <addr-line>Povo-Trento</addr-line>
          ,
          <country>Italy Potsdam</country>
          ,
          <institution>Germany University of Bergamo</institution>
          ,
          <country country="IT">Italy</country>
        </aff>
      </contrib-group>
      <abstract>
        <p>English. This paper presents the first release of LICO, a Lexicon for Italian COnnectives. LICO includes about 170 discourse connectives used in Italian, together with their orthographical variants, part of speech(es), semantic relation(s) (according to the Penn Discourse Treebank relation catalogue), and a number of usage examples.</p>
      </abstract>
    </article-meta>
  </front>
  <body>
    <sec id="sec-1">
      <title>1 Introduction</title>
      <p>Discourse connectives are explicit lexical markers
that are used to express functional relations
between parts of the discourse. As an example, the
italian word “quando” in the sentence “Quando si
preme sul bottone, la porta si apre da sola” (When
you press the button, the door opens by itself)
expresses a conditional relation between two parts of
the sentence (from now on, arguments).</p>
      <p>
        Work on discourse connectives in
Computational Linguistics was initially part of Rhetorical
Structure Theory
        <xref ref-type="bibr" rid="ref7">(Mann and Thompson, 1988)</xref>
        ,
where the focus is on discourse relations, which
are at the basis of the notion of textual coherence.
In Computational Linguistics, being able to
identify connectives is a central task in “shallow
discourse parsing”, which has become very popular
in recent years (e.g.,
        <xref ref-type="bibr" rid="ref5">(Lin et al., 2014)</xref>
        ) and
constituted the shared task of the CONLL conference
in 2015 and 20161. Downstream applications that
can benefit from shallow discourse structure are,
inter alia, sentiment analysis (e.g.,
        <xref ref-type="bibr" rid="ref3">(Bhatia et al.,
2015)</xref>
        and argumentation mining (e.g.,
        <xref ref-type="bibr" rid="ref8">(Peldszus
and Stede, 2013)</xref>
        ).
      </p>
      <p>Our work on connectives is mainly motivated
by the fact that, to the best of our knowledge, still
there is no high coverage resource of discourse
connectives available for Italian. LICO, the
Lexicon for Italian COnnectives, aims at filling this
gap, providing a repository of Italian connectives
aligned with recent developments in discourse
relations (i.e. the last version (3.0) of the Penn
Discourse Treebank (PDTB)).</p>
      <p>
        In addition, the LICO lexicon takes advantage
from DimLex, a similar repository for German
        <xref ref-type="bibr" rid="ref11 ref15 ref17 ref18">(Scheffler and Stede, 2016; Stede and Umbach,
1998)</xref>
        ; in fact DimLex served as the main
inspiration for creating LICO (see section 4).
DimLex is an XML-encoded resource that can be
used for NLP; the public version provides
information on orthographical variants, syntactic
behavior, semantic relations (in terms of PDTB),
and usage examples. It is used for automatic
discourse parsing, and also for semi-automatic
text annotation using the ConAno tool
        <xref ref-type="bibr" rid="ref16">(Stede and
Heintze, 2004)</xref>
        . Another relevant resource for
connectives is LEXCONN, for French,
        <xref ref-type="bibr" rid="ref13">(Roze et
al., 2012)</xref>
        , which contains about 300 connectives
with their syntactic category and coherence
relations from Segmented Discourse Representation
Theory
        <xref ref-type="bibr" rid="ref1">(Asher and Lascarides, 2003)</xref>
        (and to some
extent Rhetorical Structure Theory
        <xref ref-type="bibr" rid="ref7">(Mann and
Thompson, 1988)</xref>
        ).
      </p>
      <p>LICO is freely distributed under a CC-BY
licence.</p>
      <p>1http://www.cs.brandeis.edu/ clp/conll16st/</p>
    </sec>
    <sec id="sec-2">
      <title>Discourse Connectives</title>
      <p>The definition of discourse connective is
controversial both in traditional grammar and in the
linguistic literature. Our definition is based on the
encyclopedia entry on connectives by Ferrari (2010),
included in the reference work for the Italian
language recently published by Treccani. In this
entry, connectives are defined as “each of the
invariable forms [...], that introduce relations that
structure “logically” the meanings of the sentence and
of the text”2. The definition provided in Ferrari
(2010) is restrictive, as it does not include
variable forms, i.e. those forms which are subject
to morphological modifications, such as ne
consegue/conseguiva che ‘it follows/followed/ that’,
nor does it include pragmatic uses of connectives
(also known as discourse markers) such as causal
perche´ ‘why’ in “Che ore sono? Perche´ ho
dimenticato l’orologio” (‘what time is it? Because I
forgot my watch’). On the other end, it assumes that
logical relations marked by connectives hold
between events or assertions, and therefore includes
as arguments for the relation nominal expressions
such as “dopo il pressante invito ...” ‘after the
pressing invitation ...’, i.e. expressions that
contain an event nominal, - although the event is, in
this case, referred to instead of predicated.</p>
      <p>In our work, we partly drop the invariability
criteria; we do not include forms which exhibit
morphological inflection or conjugation, but we do
include connectives which show a certain degree of
lexical variability that is, multi-word expressions
which are not totally rigid from a lexical point of
view (ad esempio/per esempio ‘for example’; see
section 3).
3</p>
    </sec>
    <sec id="sec-3">
      <title>The Structure of the Lexicon</title>
      <p>Each entry in the LICO lexicon corresponds to a
connective (including its variants). Currently, for
each entry LICO specifies:
whether the connective (or its variants) is
composed by a single token (“part = single”,
e.g. perche´) or by more than one token (“part
= phrasal” e.g. di conseguenza);
whether the connective is composed by
correlating part (“orth = discont”) or not (“orth
2“Il termine connettivo indica in linguistica ciascuna delle
forme invariabili [...], che indicano relazioni che strutturano
‘logicamente’ i significati della frase e del testo”.
= cont”) and the specification of the two
correlating parts, e.g. “orth = discont”: da una
parte (“part = phrasal”), dall’altra (“part =
phrasal”); “orth = cont”: perche´ (“part =
single”);
possible orthographic variants: e.g. cio`
nonostante (“part = phrasal”) and
ciononostante (“part = single”);
possible lexical variants: e.g dopo di che´ and
dopo di cio`. Notice that in some cases this
lexical variants determine a different
syntactic environment, such as in modo da and in
modo che, the first being followed by
infinitive form, the following by a subjunctive
form;
pos category: adverbs, preposition
subordinating or coordinating conjunctions;
the semantic relation(s) that the connective
indicates, according to the PDTB 3.0 schema
(see section 3.1);
examples of the connectives for each
semantic relation;
possible alignments with lexicon of
connectives in other languages.</p>
      <p>Table 1 shows the entry for quando, which
presents more than one semantic relation, and the
entry for ciononostante, cio` nonostante,
nonostante cio`, as example of a connective with
orthografic variants in LICO.
3.1</p>
      <sec id="sec-3-1">
        <title>Semantic relations</title>
        <p>
          For the annotation of the semantic relation we
used the PDTB 3.0 schema of relations
          <xref ref-type="bibr" rid="ref11 ref18">(Webber
et al., 2016; Rehbein et al., 2016)</xref>
          as proposed in
the DimLex resource
          <xref ref-type="bibr" rid="ref11 ref15 ref18">(Scheffler and Stede, 2016)</xref>
          ,
which is our main reference resource.
        </p>
        <p>
          The schema is a most recent version of PDTB
2.0
          <xref ref-type="bibr" rid="ref10 ref9">(Prasad et al., 2008; Prasad et al., 2007)</xref>
          and
includes semantic relations structured in a
hierarchy composed by three levels. In the first level, the
class level, the relations are grouped in four major
classes: TEMPORAL, CONTINGENCY,
COMPARISON and EXPANSION. The second level, the
type level, specifies further the semantics of the
class level. For example, the TEMPORAL:
Synchronous tag is used for connectives that indicate
that the two arguments are simultaneous, while the
entry-id
orth
part
146
cont
single
quando
POS subordinating
sem relation TEMPORAL: Synchronous
ex.: Quando lascio` l’appartamento, arrivo` la chiamata
rel. to German id: 5
sem relation CONTINGENCY:Condition
ex.: Quando si preme sul bottone, la porta si apre da sola.
ex.: Quando me lo chiedi, lo lascero` stare.
        </p>
        <p>rel. to German id: 116
entry-id
orth
part</p>
        <p>variant
orth
part</p>
        <p>variant
orth
part
variant
TEMPORAL: Asynchronous tag is used for
connectives that indicate a before-after relation
between the arguments. The third level (subtype
level)3 varies according to the role of the two
arguments involved in the relation. For example,
CONTINGENCY:Cause:Reason is used if the argument
introduced by the connective -Arg2- is the
reason for the situation in the other argument
-Arg1(e.g. I stayed at home, because it was raining),
while CONTINGENCY:Cause:Results is used if
Arg2 represents the result/effect of Arg1 (e.g. It
was raining, therefore I stayed at home). Not
every type has a further subtype.</p>
        <p>In the LICO structure, each connective is assigned
with one or more three-level tags.
4</p>
      </sec>
    </sec>
    <sec id="sec-4">
      <title>The Current Resource</title>
      <p>In this Section, we present the current resource
and its construction. In particular, we focus
on describing how the list of entries has been
identified so far and how we proceeded to acquire
the semantic information for each entry.</p>
      <p>List of connectives. Currently, LICO is
composed by 173 entries, each one corresponding to
3The names of the levels are taken form Prasad et al.
(2007).
a connective and its orthographical or lexical
variants. In order to compile this list we used a
number of grammatical and lexical resources for
Italian and for other languages.</p>
      <p>
        First, we retrieved the list of connectives
mentioned by Ferrari (2010) in the Enciclopedia
Treccani for the entry connettivi4 for a total of 33
connectives. Then, we retrieved the list of connectives
tagged as congiunzione testuale in Sabatini Coletti
2006
        <xref ref-type="bibr" rid="ref14">(Sabatini-Coletti, 2005)</xref>
        discarding the ones
of literary use, for a total of 70 entries. Finally,
we benefited from the DimLex resource for
German, as we enriched our list by identifying the
equivalent Italian terms of the German
connectives5. This process was facilitated by the presence
of examples in the German resource in which the
connective is displayed in context: only the Italian
candidates that maintain the sense of the German
connectives were added to LICO. We keep trace
of this “German-Italian” links and we will use this
information to enrich also the characteristic of the
entry in LICO (e.g. aber ! ma). A total of 127
entries were collected with this method. Figure
1 shows the overlap between the three resources
and Table 2 shows a sample of the connectives in
LICO and the respective sources.
      </p>
      <sec id="sec-4-1">
        <title>Semantic relations in LICO.</title>
        <p>In LICO connectives are tagged with the
semantic relations that the connective can indicate
in a text, selecting the most appropriate ones in
the PDTB 3.0 schema. In this process we took
advantage from the information which was already
4http://www.treccani.it/enciclopedia/connettivi
(Enciclopediadell’Italiano)=, last access July 21st 2016.</p>
        <p>5https://github.com/discourse-lab/dimlex
dopo
dopo di che
dopodiche´
dopotutto
dunque
e
ebbene
eccetto
eppure</p>
        <sec id="sec-4-1-1">
          <title>Ferrari</title>
        </sec>
        <sec id="sec-4-1-2">
          <title>Treccani</title>
          <p>dopo
dunque
e</p>
          <p>
            DimLex
(equivalent)
dopo
dopodiche´ dopo di che
dopotutto
dunque
ebbene
eppure
dunque
e
eccetto
eppure
present in the resources we used for building the
list. In fact, the DimLex resource provides this
information for the German connectives, and both
the Italian resources previously mentioned
provide useful information about the semantic
relation triggered by the connective.6 A total of 23
different PTDB relations have been used to describe
LICO entries. In order to validate the tagging of
semantic relations, we conducted a research by
observing examples of the use of the connectives in
corpora, i.e. we wanted to verify whether the
relation that a connective introduces in a portion of
text is one of the relations already tagged for that
same connective in the first step. In particular, we
searched for 20 connectives in the ItWac corpus
            <xref ref-type="bibr" rid="ref2">(Baroni et al., 2009)</xref>
            and we retrieved occurrences
with 400 characters on both sides of the
connective. We limited our observation to 5 retrieved
segments of text in which the connective is actually
playing such a role. We finally tagged each
connective in each portion of text with the semantic
relation it indicates.
          </p>
          <p>To further confirm the corpus-driven evidences
for the semantic relations, we asked two
annotators (one being an expert annotator, the other not)
to perform the same tagging task. We then
calculated the interannotator agreement between the
two annotators adopting the Dice’s coefficient
(Ri6In particular, in the online version of Sabatini Coletti
(http://dizionari.corriere.it/dizionarioitaliano/D/dizionario
.shtml, last access July 21st 2016) the semantic relations the
connectives can trigger are described in the definition of the
connective itself, e.g. “quindi, cong. testuale: Con valore
deduttivo-conclusivo, percio`, di conseguenza, per questo
motivo, dunque”. Ferrari (2010) in the Enciclopedia Treccani
proposes a non hierarchical classification which includes the
following relations: “temporal relation” “causal relation”,
“consequence relation”, “condition relation”, “opposition
relations”.
jsbergen, 1997)7 for three configurations, one for
each level of the relation schema: class agreement,
type agreement, subtype agreement. We
considered that there was agreement if both annotators
identify exactly the same class, type, subtype
respectively. The Dice values result in 0,78 for class
agreement and 0,71 for both type agreement and
subtype agreement.</p>
          <p>Observing cases of disagreement, we can
make the following preliminary considerations.
The main cases of disagreement regard the
COMPARISON:Contrast relation (on one hand)
and the COMPARISON:Concession and
EXPANSION:Substitution relations (on the other
hand). These relations in fact appear to be the
ones that connect arguments that are in
contrast. As an example, the connective anziche´
‘rather than’ in Example (1) has been annotated
as COMPARISON:Contrast by annotator1 and as
EXPANSION:Substitution:Arg1-as-subst by
annotator2: the first enlightens the contrast between
“emissione attraverso il Tesoro” and “usare il
tradizionale sistema”, the second emphasises that
Arg2 represents the alternative to the Arg1.
(1)
[..] chiedeva l’ emissione di dollari in
banconote statunitensi attraverso il Tesoro
anziche´ usando il tradizionale sistema
della Federal Reserve.</p>
          <p>Another interesting case concerns the
disagreement between the relations
TEMPORAL:Asynchronous:precedence (in
which Arg2 follows Arg1) and
CONTINGENCY:Cause:Result (in which Arg2 is the
results of Arg1), being the two strictly connected
(i.e. in a cause-effect relation, the effect follows
the cause). As an example, in (2) one
annotator marks the connective as indicator of the
temporal sequence of Arg1 and Arg2, while the
other prefers to mark it as an indicator of the
cause-effect relation.
(2)
[..] Il bello e` che i tipi hanno pure
accennato a prendersela con me, al che io gli ho
abbaiato contro una sequela di insulti [..]
In general, the relations that were initially
as7Dice’s coefficient measures how similar two sets are by
dividing the number of shared elements of the two sets by
the total number of elements they are composed by. This
produces a value from 1, if both sets share all elements, to 0,
if they have no element in common.
signed to these connectives were confirmed by the
corpus-based exercise (i.e. at least one annotator
assigns the tag in at least one portions of text);
viceversa, in some cases one of the two
annotators assigned a relation that was not initially
identified.8
5</p>
        </sec>
      </sec>
    </sec>
    <sec id="sec-5">
      <title>Conclusion and Further work</title>
      <p>In this paper we have presented LICO, a new
resource for the Italian language describing lexical
properties of discourse connectives. While LICO
fills a gap with respect to similar resources
existing for other languages, it is still under
construction under several aspects. Our short term plans
include the completion of the lexical entries with
corpus derived examples and the observation of
the connectives in Italian corpora, in order to
acquire more information about the semantic
relations that each connective can indicate and thus
extend the annotation of the semantic relations in
LICO.</p>
    </sec>
    <sec id="sec-6">
      <title>Acknowledgment</title>
      <p>We acknowledge Denise Pangrazzi for her
contribution to identify the Italian equivalents of the
German connectives.</p>
    </sec>
  </body>
  <back>
    <ref-list>
      <ref id="ref1">
        <mixed-citation>
          <string-name>
            <given-names>Nicholas</given-names>
            <surname>Asher</surname>
          </string-name>
          and
          <string-name>
            <given-names>Alex</given-names>
            <surname>Lascarides</surname>
          </string-name>
          .
          <year>2003</year>
          .
          <article-title>Logics of conversation</article-title>
          . Cambridge University Press.
        </mixed-citation>
      </ref>
      <ref id="ref2">
        <mixed-citation>
          <string-name>
            <given-names>Marco</given-names>
            <surname>Baroni</surname>
          </string-name>
          , Silvia Bernardini, Adriano Ferraresi, and
          <string-name>
            <given-names>Eros</given-names>
            <surname>Zanchetta</surname>
          </string-name>
          .
          <year>2009</year>
          .
          <article-title>The wacky wide web: a collection of very large linguistically processed web-crawled corpora</article-title>
          .
          <source>Language resources and evaluation</source>
          ,
          <volume>43</volume>
          (
          <issue>3</issue>
          ):
          <fpage>209</fpage>
          -
          <lpage>226</lpage>
          .
        </mixed-citation>
      </ref>
      <ref id="ref3">
        <mixed-citation>
          <string-name>
            <given-names>Parminder</given-names>
            <surname>Bhatia</surname>
          </string-name>
          , Yangfeng Ji, and
          <string-name>
            <given-names>Jacob</given-names>
            <surname>Eisenstein</surname>
          </string-name>
          .
          <year>2015</year>
          .
          <article-title>Better document-level sentiment analysis from rst discourse parsing</article-title>
          .
          <source>In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing</source>
          , Lisbon, Portugal, September. Association for Computational Linguistics.
        </mixed-citation>
      </ref>
      <ref id="ref4">
        <mixed-citation>
          <string-name>
            <given-names>Angela</given-names>
            <surname>Ferrari</surname>
          </string-name>
          .
          <year>2010</year>
          .
          <article-title>Connettivi. In Enciclopedia dell'Italiano. diretta da Raffaele Simone, con la collaborazione di Gaetano Berruto e Paolo D'Achille</article-title>
          , Roma, Istituto della Enciclopedia Italiana.
        </mixed-citation>
      </ref>
      <ref id="ref5">
        <mixed-citation>
          <string-name>
            <given-names>Ziheng</given-names>
            <surname>Lin</surname>
          </string-name>
          , Hwee Tou Ng, and
          <string-name>
            <surname>Min-Yen Kan</surname>
          </string-name>
          .
          <year>2014</year>
          .
          <article-title>A pdtb-styled end-to-end discourse parser</article-title>
          .
          <source>Natural Language Engineering</source>
          ,
          <volume>20</volume>
          :
          <fpage>151</fpage>
          -
          <lpage>184</lpage>
          .
        </mixed-citation>
      </ref>
      <ref id="ref6">
        <mixed-citation>
          <article-title>8For this moment, the “new” relations are not included in LICO.</article-title>
        </mixed-citation>
      </ref>
      <ref id="ref7">
        <mixed-citation>
          <string-name>
            <surname>William C Mann and Sandra A Thompson</surname>
          </string-name>
          .
          <year>1988</year>
          .
          <article-title>Rhetorical structure theory: Toward a functional theory of text organization</article-title>
          .
          <source>Text-Interdisciplinary Journal for the Study of Discourse</source>
          ,
          <volume>8</volume>
          (
          <issue>3</issue>
          ):
          <fpage>243</fpage>
          -
          <lpage>281</lpage>
          .
        </mixed-citation>
      </ref>
      <ref id="ref8">
        <mixed-citation>
          <string-name>
            <given-names>Andreas</given-names>
            <surname>Peldszus</surname>
          </string-name>
          and
          <string-name>
            <given-names>Manfred</given-names>
            <surname>Stede</surname>
          </string-name>
          .
          <year>2013</year>
          .
          <article-title>From argument diagrams to argumentation mining in texts: A survey</article-title>
          .
          <source>International Journal of Cognitive Informatics and Natural Intelligence (IJCINI)</source>
          ,
          <volume>7</volume>
          (
          <issue>1</issue>
          ):
          <fpage>1</fpage>
          -
          <lpage>31</lpage>
          .
        </mixed-citation>
      </ref>
      <ref id="ref9">
        <mixed-citation>
          <string-name>
            <given-names>Rashmi</given-names>
            <surname>Prasad</surname>
          </string-name>
          , Eleni Miltsakaki, Nikhil Dinesh,
          <string-name>
            <given-names>Alan</given-names>
            <surname>Lee</surname>
          </string-name>
          , Aravind Joshi, Livio Robaldo, and Bonnie L Webber.
          <year>2007</year>
          .
          <article-title>The Penn Discourse Treebank 2.0 Annotation Manual</article-title>
          .
        </mixed-citation>
      </ref>
      <ref id="ref10">
        <mixed-citation>
          <string-name>
            <given-names>Rashmi</given-names>
            <surname>Prasad</surname>
          </string-name>
          , Nikhil Dinesh,
          <string-name>
            <given-names>Alan</given-names>
            <surname>Lee</surname>
          </string-name>
          , Eleni Miltsakaki, Livio Robaldo,
          <article-title>Aravind K Joshi,</article-title>
          and Bonnie L Webber.
          <year>2008</year>
          .
          <article-title>The Penn Discourse TreeBank 2.0</article-title>
          .
          <source>In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC'08)</source>
          , Marrakech, Morocco, May.
        </mixed-citation>
      </ref>
      <ref id="ref11">
        <mixed-citation>
          <string-name>
            <given-names>Ines</given-names>
            <surname>Rehbein</surname>
          </string-name>
          , Merel Scholman, and
          <string-name>
            <given-names>Vera</given-names>
            <surname>Demberg</surname>
          </string-name>
          .
          <year>2016</year>
          .
          <article-title>Annotating Discourse Relations in Spoken Language: A Comparison of the PDTB and CCR Frameworks</article-title>
          .
          <source>In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC</source>
          <year>2016</year>
          ), Portorozˇ, Slovenia, May.
        </mixed-citation>
      </ref>
      <ref id="ref12">
        <mixed-citation>
          <string-name>
            <surname>CJ van Rijsbergen</surname>
          </string-name>
          .
          <year>1997</year>
          . Information retrieval.
          <year>1979</year>
          .
        </mixed-citation>
      </ref>
      <ref id="ref13">
        <mixed-citation>
          <string-name>
            <given-names>Charlotte</given-names>
            <surname>Roze</surname>
          </string-name>
          , Laurence Danlos, and
          <string-name>
            <given-names>Philippe</given-names>
            <surname>Muller</surname>
          </string-name>
          .
          <year>2012</year>
          .
          <article-title>LEXCONN: a French lexicon of discourse connectives</article-title>
          .
          <source>Discours</source>
          . Revue de linguistique, psycholinguistique et informatique., (
          <volume>10</volume>
          ).
        </mixed-citation>
      </ref>
      <ref id="ref14">
        <mixed-citation>
          <string-name>
            <given-names>Il</given-names>
            <surname>Sabatini-Coletti</surname>
          </string-name>
          .
          <year>2005</year>
          .
          <article-title>Dizionario della lingua italiana 2006, con CD-ROM</article-title>
          . Milano, Rizzoli Larousse.
        </mixed-citation>
      </ref>
      <ref id="ref15">
        <mixed-citation>
          <string-name>
            <given-names>Tatjana</given-names>
            <surname>Scheffler</surname>
          </string-name>
          and
          <string-name>
            <given-names>Manfred</given-names>
            <surname>Stede</surname>
          </string-name>
          .
          <year>2016</year>
          .
          <article-title>Adding Semantic Relations to a Large-Coverage Connective Lexicon of German</article-title>
          .
          <source>In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC</source>
          <year>2016</year>
          ), Portorozˇ, Slovenia, May.
        </mixed-citation>
      </ref>
      <ref id="ref16">
        <mixed-citation>
          <string-name>
            <given-names>Manfred</given-names>
            <surname>Stede</surname>
          </string-name>
          and
          <string-name>
            <given-names>Silvan</given-names>
            <surname>Heintze</surname>
          </string-name>
          .
          <year>2004</year>
          .
          <article-title>Machineassisted rhetorical structure annotation</article-title>
          .
          <source>In Proceedings of the 20th International Conference on Computational Linguistics</source>
          , pages
          <fpage>425</fpage>
          -
          <lpage>431</lpage>
          , Geneva.
        </mixed-citation>
      </ref>
      <ref id="ref17">
        <mixed-citation>
          <string-name>
            <given-names>Manfred</given-names>
            <surname>Stede</surname>
          </string-name>
          and
          <string-name>
            <given-names>Carla</given-names>
            <surname>Umbach</surname>
          </string-name>
          .
          <year>1998</year>
          .
          <article-title>Dimlex: A lexicon of discourse markers for text generation and understanding</article-title>
          .
          <source>In Proceedings of the 17th international conference on Computational linguisticsVolume 2</source>
          , pages
          <fpage>1238</fpage>
          -
          <lpage>1242</lpage>
          . Association for Computational Linguistics.
        </mixed-citation>
      </ref>
      <ref id="ref18">
        <mixed-citation>
          <string-name>
            <given-names>Bonnie</given-names>
            <surname>Webber</surname>
          </string-name>
          , Rashmi Prasad,
          <string-name>
            <given-names>Alan</given-names>
            <surname>Lee</surname>
          </string-name>
          ,
          <string-name>
            <given-names>and Aravind</given-names>
            <surname>Joshi</surname>
          </string-name>
          .
          <year>2016</year>
          .
          <article-title>A Discourse-Annotated Corpus of Conjoined VPs</article-title>
          .
          <source>In Proceedings of the 10th Linguistic Annotation Workshop held in conjunction with ACL</source>
          <year>2016</year>
          (
          <article-title>LAW-X 2016)</article-title>
          , pages
          <fpage>22</fpage>
          -
          <lpage>31</lpage>
          . Association for Computational Linguistics.
        </mixed-citation>
      </ref>
    </ref-list>
  </back>
</article>