ArticlePDF Available

Representing temporal discourse markers for generation purposes Brigitte Grote



Discourse markers are an important means to signal the kind of coherence relation holding between adjacent text spans. Research on generating discourse markers has been mainly concerned with causal markers, whereas temporal markers have not received much attention. In this paper, we identify semantic, pragmatic and syntactic features that axe required to support a motivated choice of German temporal subordinating conjunctions and prepositions during text production. Information on individual markers is assembled in a discourse marker lexicon, which is used as a declaxative resource at the sentence planning stage. We illustrate how this resource can be used to produce alternative verbalizations of the temporal relationship holding between two events.
Representing temporal discourse markers for generation purposes
Brigitte Grote
Otto-von-Guericke Universit~it Magdeburg
Institut fiir Wissens- und Sprachverarbeitung
P.O. Box 4120, 39016 Magdeburg, Germany
emaih grote~iws, cs. uni-magdeburg, de
Discourse markers are an important means to signal
the kind of coherence relation holding between adja-
cent text spans. Research on generating discourse
markers has been mainly concerned with causal
markers, whereas temporal markers have not received
much attention. In this paper, we identify semantic,
pragmatic and syntactic features that are required
to support a motivated choice of German temporal
subordinating conjunctions and prepositions during
text production. Information on individual markers
is assembled in a discourse marker lexicon, which is
used as a declarative resource at the sentence plan-
ning stage. We illustrate how this resource can be
used to produce alternative verbalizations of the tem-
poral relationship holding between two events.
1 Motivation
In text, discourse markers signal the kind of coher-
ence relation holding between adjacent text spans.
For any but the most trivial applications of language
generation, motivated marker choice is an important
task. Whereas several studies have been concerned
with causal markers and their interactions with other
linguistic means, for instance, Vander Linden and
Martin (1995), Rhsner and Stede (1992), Delin et al.
(1996), temporal markers as signals of the temporal
relation holding between two events have not received
much attention, with the exception being Dorr and
Gaasterland (1995). However, quite often, it is only
by means of explicit temporal markers that the cor-
rect interpretation of a text can be ensured, as the
following examples illustrate:
(1) Der
Abstand ist nachzumessen, nachdem das Band
mindestens einen Umlauf ausge~hrt
hat. (Check
the distance after the belt has completed at least
one round.)
W~ihrend Sie den Toaster betreiben, die
Brotschlitze nicht abdecken.
(While operating the
toaster, do not cover the bread slots.)
In both cases, the order of events as recounted in
the text does not correspond to their order of oc-
curence: In example (1),
(after) marks the
event denoted in the second clause as temporally an-
terior to the one denoted in the first clause. In ex-
ample (2),
(while) marks both events as
cotemporal. Without a temporal marker, the order
of events would not be obvious.
Selecting an appropriate discourse marker for a
given temporal relation is by no means a straightfor-
ward matter, though. First, one and the same tem-
poral relation is verbalized differently depending on
the syntactic and lexical properties of the clauses it
conjoins. In German, the language we consider, such
properties are, among others, tense, aspect, and syn-
tactic structure. Interdependencies occur when two
temporal events are realized in the same sentence; we
will therefore restrict the study of German temporal
markers to subordinating conjunctions and preposi-
tions. Second, temporal discourse markers can even
overwrite the temporal relations indicated by other
linguistic means like tense and aspect, as noted by re-
searchers who work in the analysis of temporal mark-
ers (e.g. Hitzeman et al. 1995).
In a nutshell, when aiming at selecting an ap-
propriate temporal marker in text generation one
needs a representation of temporal markers that en-
ables marker choice and other sentence level decisions
(such as tense and aspect selection) to mutually con-
strain each other. In this paper, we take up the idea
suggested in Grote and Stede (1998) of a discourse
marker lexicon as a declarative resource at the sen-
tence planning stage. We demonstrate how such a
lexicon for temporal markers can be employed in text
generation to produce different verbalizations of the
same underlying temporal organization depending on
other generation decisions.
The paper is organized as follows: Section 2 re-
views related work on generating temporal mark-
ers. Section 3 describes the major semantic, prag-
matic and syntactic properties of German temporal
prepositions and subordinating conjunctions. Sec-
tion 4 presents the generation perspective: It briefly
discusses the shape of the discourse marker lexi-
con, introduces the features used in the lexicon, and
presents sample lexicon entries and their application
in the generation process.
2 Related work
Work on discourse marker generation in general has
focussed on marker selection, mainly for causal re-
lations (Elhadad and McKeown 1990; Vander Lin-
den and Martin 1995), and on the realization of
RST's subject-matter relations (Rhsner and Stede
1992; Delin et al. 1996). As for temporal mark-
ers, Dorr and Gaasterland (1995) examine the gen-
eration of English temporal subordinating conjunc-
tions. Gagnon and Lapalme (1993), on the other
hand, describe the generation of French temporal ad-
verbs based on a DRT representation of the discourse.
While Gagnon and Lapalme (1993) only briefly
address conjunctions and prepositions, Dorr and
Gaasterland (1995) present a detailed study of tem-
poral connectives, but they consider English mark-
ers only. The only account on automatically produc-
ing German temporal expressions that we know of is
Ehrich (1987); however, she discusses the interaction
of tense and aspect in simple sentences only.
Most studies that deal with discourse markers re-
gard their production as a mere consequence of other
sentence level decisions such as aggregation, lexical-
ization, syntactic structuring, and--in the case of
temporal markers--as determined by tense and as-
pectual choices. We believe, however, that one needs
a more flexible control to increase the expressiveness
of generation systems. Although there have been
quite a few studies on individual aspects of sentence
planning, little attention has been paid to the in-
teraction between the various tasks--exceptions are
Rambow and Korelsky (1992) and Wanner and Hovy
(1996)--and in particular to the role of marker choice
in the overall sentence planning process.
There exists a large body of research in NLU on
analysing the temporal structure of texts, including
the role of temporal markers, though again restricted
to English (Moens and Steedman 1988; Lascarides
and Oberlander 1993; Hitzeman et al. 1995). We turn
to these studies when it comes to identifying the in-
formation that needs to be assembled for representing
temporal markers.
3 Linguistic perspective:
Describing temporal markers
Selecting an appropriate German temporal marker
given two events in a temporal relationship requires
detailed knowledge of the semantic, pragmatic and
syntactic properties that characterize temporal mark-
ers. This section introduces the major properties and
the correlations between temporal markers
and other linguistic means that indicate temporal or-
ganization. We base our account on two sources:
descriptive linguistic studies, mainly by Helbig and
Buscha (1991), B/iuerle (1995), Buscha (1989) and
Steube (1980); and our analysis of temporal marker
usage in the German LIMAS corpus (Glas 1975).
3.1 The 'meaning' of German temporal
Temporal subordinating conjunctions and temporal
prepositions conjoin two events where the event in the
subordinate clause (or the PP) provides the tempo-
ral framework for interpreting the event in the main
clause: Bevor Sie den Toaster reinigen, den Net-
zstecker ziehen. (Before you clean the toaster, un-
plug the device.) and the corresponding 'shorthand'
form l br dem Reinigen des Toasters den Netzstecker
ziehen (Unplug before cleaning the toaster).
Semantic properties German grammars such as
Helbig and Buscha (1991) list about 20 temporal
subordinating conjunctions and 20 temporal prepo-
sitions. Their semantics is usually described by the
kind of temporal relation they establish between two
events, see for instance, Steube (1980) and Helbig and
Buscha (1991): The event in the main clause can ei-
ther overlap with (simultaneity), succeed (anteri-
ority), or precede (posteriority) the event depicted
in the subordinate clause or the prepositional phrase.
In table 1 we provide a synthesis of the classifica-
tions of the most frequent German temporal mark-
ers by Helbig and Buscha (1991), Buscha (1989) and
B/iuerle (1995). The markers listed in the table re-
flect the scope of the marker study in this paper.
Two aspects are especially prominent: First, each
of the three temporal relations can be realized by a
number of temporal markers. Alternatives within a
class differ in that they realize some additional mean-
ing aspect. Consider the markers of simultaneity:
Solange, for instance, conveys the idea of a strict si-
multaneity where two events have the same start and
end time, and is more specific than w~hrend; sooft,
to give another example, highlights the concurrence
of two events.
Second, table 1 shows that some markers are am-
biguous: Als and wenn occur in all three classes, seit-
dem, sobald and sooft in two. Apparently, neither
of them has any special temporal implicature on its
own; instead, these markers depend on syntactic and
lexical contexts to receive an umambiguous temporal
meaning. We will return to this issue in section 3.2.
Pragmatic properties The choice Of a particular
marker to express a temporal relation between two
events interacts with the focus structure as in:
(3) (a) Bevor ihr Mann das Haus verlie[J, ging sie zur
Arbeit. (Before her husband left the house, she
went to work.)
(b) Nachdem sie zur Arbeit gegangen war, verliefl
ihr ivIann das Haus. (After she had gone to work,
her husband left the house.)
Alternatives (3a) and (3b) both express that the
event of 'going to work' precedes the event of 'leav-
ing the house'. They differ in that they focus on
temporal relation temporal markers
simultaneity subc: als
(as), indes(sen) (meanwhile), seitdem (since), sobald (as
soon as), solange
(as long as), sooft (whenever),
sowie (as soon as),
w~hrend (while),
wenn (when)
(at), auf (on), bei (during),
(within), durch (for), in (in), iiber (over),
w~hrend (during)
anteriority subc: a./s
(when), kaum dab (no sooner),
nachdem (after),
seit(dem) (since),
sobald (as
soon as),
sooft (whenever), sowie (as soon as), wenn (when)
prep: ab (from), nach (after),
posteriority subc:
als (when), bevor (before), his (until), ehe (before),
prep: his (until), vor (before)
Table 1:
German temporal subordinating conjunctions (subc) and prepositions (prep) classified by temporal
relations. Note that
corresponding English markers
only approximate translations.
different aspects of the situation: In (3a) the ear-
lier event is in the centre of attention, in (3b) the
later one (assuming that the matrix sentence is more
prominent). This phenomenon interacts with other
discourse phenomena, for instance, given and new
information, and--when placed in a larger discourse
context--with presuppositions and their accommo-
dation (Lascarides and Oberlander 1993). However,
the treatment of the discourse behaviour of temporal
markers is beyond the scope of this paper.
Pragmatic issues further concern style. Regard-
ing temporal markers, stylistic features are of minor
importance: We only observe variation between ar-
chaic and neutral (da vs. als), and formal and neutral
(kaum dab vs.
3.2 Syntactic and lexical constraints
When expressing several events in the same sentence,
marker choice interacts with other linguistic means:
Temporal markers impose particular constraints on
the syntactic and lexical contexts they can occur in.
Conversely, these contexts can influence the meaning
of markers. 1
Markers and Aktionsart/aspect Aspect is tra-
ditionally taken to have two components, the non-
inherent grammatical features, and the inherent lex-
ical features. Inherent features characterize facets of
the situation denoted by a verb, for instance, whether
it is an event or a state. We will label these fea-
to avoid confusion. According to
Bussmann (1990), the major Aktionsarten in German
are stative
(wissen/to know)
and dynamic. For the
latter, the basic dichotomy is that between durative
sleep) and non-durative verbs, which are
subdivided into iterative
(flattern/to flap),
(klopfen/to knock),
burn up)
and causative verbs
1Traditional grammars, which the present account is based
on, usually list aspect, Aktionsart and tense as constraining
parameters on marker choice. However, there is no consensus
on the role of these parameters; B~.uerle (1995) provides a good
overview of the range of positions.
Two kinds of interdependencies are generally ac-
knowledged, see Ehrich (1987), Buscha (1989) and
B~iuerle (1995). First, temporal markers are sensi-
tive to the Aktionsart of a verb. Consider
and als which can both express simultaneity:
(4) (a)
Als das Kabel schmolz / riB,
ich nicht im
(When the cable melted / tore, I wasn't in
the room.)
W~hrend das
schmolz / *riB, war ich
nicht im Raum.
(While the cable melted / *tore, I
wasn't in the room.)
expects a durative verb in the subordi-
nate clause, hence it can occur with
but not with
reit3en/to tear. Als,
in contrast,
can be used with durative and resultative verbs, as
(4a) illustrates. Second, temporal markers may even
shift the Aktionsart of a verb, for instance from a
semelfactive reading to an iterative one as in:
(5) (a)
es an
der Tfir klopft, schreit das Baby.
(When someone knocks at the door, the baby cries.)
Wiihrend es an der Tfir klopft, schreit das
(While someone knocks at the door, the
baby cries.)
Grammatical aspect reflects the individual per-
spective a speaker adopts with respect to an event,
such as perfective (temporally closed) or imperfec-
tire. In German, this distinction is grammatically
realized by choosing a perfective or simple tense, e
Aktionsart and aspect closely interact, consider ex-
ample (6) where the anterior reading (6b) is due to
the use of a perfective tense with a non-durative verb
in the subordinate clause, which indicates that the
activity has been concluded:
(6) (a)
Seitdem ich ihn kenne, ist
(Since I know him, he is a non-smoker.)
Seitdem seine I~rau gestorben ist, sehe ich ihn
nur selten.
(Since his wife has died, I only rarely
see him.)
Sin contrast to English, and especially to slavic languages,
German has no elaborate aspect system: Distinctions like pro-
gressive and simple cannot be signalled by morphological fea-
tures of the verb, but require a separate temporal adverb: He
is reading
Sie 1lest gerade
(She reads right now).
Here, verb properties determine the reading of the
temporal marker. Our study of temporal marker oc-
currences in the LIMAS corpus suggests that mark-
ers belonging to the simultaneity class typically real-
ize imperfective aspect, whereas temporal connecting
words that signal anteriority correlate with a perfec-
tive aspect in the subordinate clause.
Markers and verbal tense Some markers can
only be used with particular tenses, for instance, a/s
in its simultaneous reading cannot occur with present
tense, whereas
wenn as
signal of simultaneity corre-
lates with present and past tense:
(7) (a)
in Dresden war (*ist), suchte (,sucht) er
seine 1~reundin
in Dresden ist / war, sucht / suchte er
seine Freundin
However, tempus sensitivity of temporal markers is
not a matter of the grammatical tense form (such as
simple past, present perfect, etc.) but relates to the
temporal structure of the individual events, and to
how their temporal structures are related. Assum-
ing the Reichenbachian threefold distinction between
Event Time (E), Reference Time (R), and Speaking
Time (S) (the Basic Tense Structure, BTS, (Reichen-
bach 1947)), we observe that the constraints imposed
by a marker on verb tense concern the underlying re-
lation between E and S of both clauses: Selecting
either a/s or
to express simultaneous events in
the main clause (era) and in the subordinate clause
(es) depends on whether the event times precede S
(E(em),E(es)_S) or concur with S (E(em),E(es),S). 3
The grammatical tense results from combining the
BTS of both clauses and their aspectual features.
Markers and syntactic structure The most
straightforward correlation is that between syntac-
tic structure and marker choice: If two events are
expressed by a hypotactic structure, a subordinating
conjunction is required. When a deverbal realization
of an event is possible (e.g.
Treffen; to
meet/the meeting), a clause with an adverbial (tem-
poral) prepositional phrase is realized.
Markers and temporal quantifiers With some
markers, the temporal relation denoted by the marker
can be quantified by a temporal adverb as in
(shortly before) or
einige Stunden nachdem
(several hours after); others cannot be quantified:
einige Stunden sobald
(.several hours as soon as).
4 Generation perspective:
Representing temporal markers
A representation of temporal markers suitable for
generation purposes has to accommodate the follow-
ing demands: First, it has to describe the semantic
SThe comma stands for 'is cotemporal', the underscore for
and pragmatic features of markers in a manner that
supports a motivated choice between markers which
can realize the same temporal relation. Second, it
has to account for the constraints temporal mark-
ers impose on their syntactic and lexical contexts,
thereby enabling interactions between marker choice
and other sentence planning decisions where the order
of decision-making is not fixed. In Grote and Stede
(1998) we argue that such a flexible control is best
realized by introducing independent modules for the
different sentence planning tasks, such as proposed
by Wanner and Hovy (1996), and that these modules
should rely on declarative representations as much as
possible. Therefore, we propose a discourse marker
lexicon, i.e. an independent lexical resource that as-
sembles specifically the information associated with
discourse markers.
Traditional lexicology and grammars describe lexi-
cal entries along three features: semantic, pragmatic
and syntactic dimensions (see section 3). From the
production perspective, these features are to be clas-
sifted with respect to when and where they come into
play in the generation process; this amounts to a pro-
cedural view on the information coded in the lexicon.
Following Grote and Stede (1998) we assume three
categories in the marker lexion:
Applicability conditions: The necessary con-
ditions that need to be present in the input rep-
resentation for the marker to be a candidate.
Chiefly, this is the semantic/discourse relation
to be expressed, and also (if applicable) features
pertaining to presuppositions and intentions.
Combinability conditions: The constraints
that the marker imposes on its neighbouring lin-
guistic constituents (the 'syntagmatic' dimen-
sion). These are syntactic constraints on subcat-
egorization and semantic type constraints, which
interact with other realization decisions in sen-
tence planning.
Distinguishing features: If preferential choice
dimensions, such as style, brevity, etc., are at-
tended to in the system, then these features serve
to distinguish markers that are otherwise equiv-
alent (the 'paradigmatic' dimension).
In the remainder of this section we describe lexicon
entries for temporal markers along these lines.
4.1 Applicability conditions
Semantic conditions The semantic classes intro-
duced in section 3.1 (simultaneity, anteriority and
posteriority) turned out to be too coarse for genera-
tion purposes. Instead, one needs a more fine-grained
representation of the semantics of temporal markers
to support an informed choice among markers within
the broad classes.
Allen's temporal interval relationships provide an
adequate framework (Allen 1984), as already sug-
gested by Dorr and Gaasterland (1995). Allen in-
troduces seven basic temporal interval relationships,
equals(=), after(>), during(d), overlaps(o),
meets(m), starts(s), finishes(f)--and
their inverses
<,di, oi, mi, si,fi--that
may exist between two events
and es. For instance,
overlaps(em,e~) as
in (4b)
implies that there is an intersection between the time
at which em occurs and the time at which
but that neither event is a subset of the other.
Each temporal relation corresponds to one or sev-
eral German temporal markers, for instance,
may be expressed by the entire range of simultane-
ity markers given in table 1, except for solange and
kaum daB. Conversely, the majority of the tempo-
ral markers can realize several temporal interval re-
lations. Take the connective
nachdem as
in example
(1), which can have the following meanings,
after(ern, es)
meets-i(em, es)
w~ihrend as
in example (4b),
equals(ern, es) A during(era, es) A starts(era, es)
A finishes(era, es) A overlaps-i(em, es)
whereas solange has only one reading:
equals(era, es ).
This adequately captures the semantic difference
In the lexicon, the ap-
plicability conditions of a particular temporal marker
are now described by listing the temporal interval re-
lations it can realize.
conditions In section 3.1 we briefly
discussed pragmatic features of temporal markers.
For the time being, the lexicon supports the features
style, with the values neutral, brief, formal,
archaic, and intention. Its value evaluative in-
dicates the speaker's (negative) attitude towards the
kind of temporal relation holding between two events
(Steube 1980; Buscha 1989).
4.2 Combinability conditions
Combinability conditions appear as constraints in the
lexicon entries of individual markers. In the present
lexicon, constraints are described using the following
Aktionsart The Aktionsart plays a central role
during the lexicalization of events: Candidate verbs
are, among others, selected due to their Aktion-
sart. Aktionsart features are usually stored in the
lexicon entries of verbs, and are thus available to
sentence planning. To represent these constraints,
we turn to Bussmann (1990) for the major Aktion-
sarten in German (see also section 3.2). 4 At present,
the lexicon supports a subset of Bussmann's Aktion-
sarten, namely stative, durative, iterative,
semelfactive, causative and resultative.
4There is no
generally accepted and
well-defined set of Ak-
tionsart features; we
opted for
Bussmann (1990) because these
features are supported by
the lexicalization component we in-
tend to use (Stede 1996).
Aspect Grammatical aspect is encoded using the
feature values perfective and imperfective.
We argued above that marker choice relates
to the underlying temporal structure---as expressed
in terms of the Reichenbachian threefold description
of time-and not to a particular grammaticai tense
(see also Ehrich (1987)). Temporal constraints in the
marker lexicon will thus be described using the BTS
notation, and defining the legal linear orderings of
E, R and S of the related events. For instance,
in its simultaneous meaning imposes the constraint
which can be realized by all gram-
matical tenses that meet this constraint.
Mapping this representation into grammatical
tense requires knowledge on how to map pairs of Ba-
sic Tense Structures to the tense structure of complex
German sentences, as described in Hornstein (1990)
for English (Complex Tense Structures, CTS) and ex-
tended by Dorr and Gaasterland (1995) to cover in-
tervals, too. Since we envision independent modules
for the different sentence planning tasks that posit
their choices as constraints, the tense selection pro-
cess need not concern us.
Syntactic structure Possible values are pp
(prepositional phrase) and subord (subordinate
clause); both refer to the realization of the event that
acts as temporal reference point.
Quantification The lexicon contains the two val-
quantifiable and not-quantifiable.
4.3 The shape of the lexicon
The possible values for the applicability and combin-
ability features can now be used in the lexicon to
describe individual temporal markers. Table 2 gives
the lexical representations for most of the German
anteriority markers and the posteriority marker be-
Similar representations have been developed
for the other marker classes given in table ]. si-
multaneity and posteriority. Feature values for in-
dividual markers have been identified by analysing
marker occurrences in the UMAS corpus (Glas 1975);
as such, they mainly reflect marker
We then
compared our marker descriptions to results from re-
search literature (see section 3). Note that combin-
ability conditions can apply to main and subordinate
clause/prepositional phrase separately, hence some
feature values are prefixed with me:, sc: and
to mark their scope. If a marker involves no con-
straint for a particular feature, the slot in the table
remains empty.
Table 2 contains an informal description of the lexi-
con entries; the formal representation depends on the
actual sentence planner used in text production, see
Grote and Stede (1998) for a preliminary proposal.
nachdem (after)
sobald (as
soon as) kaum dab (no sooner bevor (before)
- denotation after(e,,, e~)A after(e,,, e,)A meets-i(e,~, e~) meets-i(e,~, e, ) before(e~, e~)A
meets-i(em, e,) meets-i(e~, e~) meets(e~, e,)
- tense
E(e~)...E(e~), SA
{E, S}(e~)
= {E, S}(e,)A
E(e,), E(e,,)_S
{E, S}(e~)
= {E, S}(e,)
- syntax pp subord subord
- quantifier quantifiable I quantifiable not-quantifiable not-quantifiable quantifiable
- style neutral brief neutral formal neutral
- intention evaluative
Table 2: Lexicon entries for some German temporal markers
4.4 Selecting temporal markers
This section briefly addresses the issue of selecting an
appropriate temporal marker during text production
using the discourse marker lexicon. We will focus on
the anteriority markers.
In our scenario, generation starts from a conceptual
representation which contains the facts that must be
reported in the text and their position in time. Let
us assume the following very simple input structure: 5
arrive (he, home, 19 : 14)
watch(he ,TV, 19 : 15,22 : 30)
The first event precedes the second event, but they
'meet' at one point in time. Now, the first step is
to determine the applicable temporal relations. Two
interpretations are possible, depending on the dis-
course context and focus structure, which we have
not dealt with so far: Focussing on the earlier event
would yield the temporal relation meets(era, es), with
em = el, focussing on the later event the relation
meets-i(e,n, e,), with e,n = e2. Matching this against
the lexicon entries in table 2 would produce bevor in
the former case, and nachdem, nach, sobald, kaum
dat3 as candidate realizations for the latter interpre-
tation. Possible verbalizations are:
(a) Bevor
Fernsehen Euckte, ist
er navh Hause
(Before he watched TV, he has come home.)
(b) Sobald er nach Hause gekommen war, guckte er
(As soon as he had
come home,
he watched TV.)
(c) (Direkt) nachdem er nach Hause gekommen
war, gucckte er Fernsehen.
is an abridged representation.
We will eventually
represent the
facts as SitSpecs (Stede 1996), which will be an-
notated with temporal information. During lexicalization--as
one task in the sentence
planning phase---SitSpecs are mapped
onto semantic representations
((Right) after he had come home, he watched TV)
(d) Nach dem Heimkommen guckte er Fernsehen.
(After coming home he watched TV.)
(e) Kaum dab er nach Hause gekommen war,
guckte er Fernsehen.
(As soon as he had come home, he watched TV.)
Assuming the anteriority interpretation (Sb-e),
how does a generation system choose among the four
remaining alternatives? We argued above that we
envision a modular architecture where independent
sentence planning modules posit their constraints re-
garding tense selection, lexicalization, syntactic real-
ization, etc. 6 In case no constraints are put forward
by the sentence planning modules, sobald (Sb) would
be selected, as it is the most specific and at the same
time neutral realization. If, however, a quantifier is
to be included, then nachdem would be chosen i~c).
If brevity is a stylistic concern, and the process in the
subordinate clause can be deverbalized, a phrasal re-
alization with the preposition nach is selected (Sd).
If, on the other hand, a more formal realization is
the overall goal given to the generator, kaum daft (Se)
would be chosen. In these cases, marker choice would
posit constraints (as given in the combinability slot
in table 2) on all other sentence planning decisions.
So far, we only considered a perfective aspect in
the subordinate clause. Once we change aspect to
imperfective, a realization including nachdem is no
longer an option, compare ,Nachdem er nach Hause
kam, hat er Fernsehen geguckt (After he came home,
he has watched TV). Sobald would be an adequate
realization. Likewise, changing the Aktionsart from
resultative to durative, as in Sobald er schlMt, guckt
6This approach differs from Dorr and Gaasterland (1995)
who impose a strict order on the selection of tense, aspect and
connecting word.
sie Fernsehen (As soon as he sleeps, she watches TV)
would rule out nachdem. With the resultative variant
einschlafen (fall asleep) both markers are possible.
Finally, if a constraint is posited that the tense has
to be 'present', kaum dab would not be available.
5 Conclusion and
Temporal markers have neither received much atten-
tion in NLG, nor has a principled account of marker
selection as such been introduced. In this paper
we presented a general framework for representing
German temporal markers for generation purposes.
We identified some of the features required to de-
scribe applicability conditions, constraints and pref-
erences, and proposed a declarative lexical resource
that makes it possible to treat temporal markers and
other linguistic means as mutual constraints at the
sentence planning stage. Now, we need to examine
individual temporal markers more closely and incor-
porate the temporal marker lexicon into a text gen-
eration system.
For the purpose Of this paper, we have assumed
that temporal relations are always explicitly sig-
nalled, and thus limited our study to marker selec-
tion. Marker occurrence, however, is an important
issue. First, Hitzeman et al. (1995) argue that there
exist temporal defaults of the kind "An event will
occur just after a preceding event"; this renders the
introduction of explicit markers superfluous. Second,
we have only assumed pairs of time-stamped expres-
sions, but have ignored that they usually occur in
a larger discourse situation where other kinds of co-
herence relations might hold between events. For in-
stance, all causal coherence relations have some tem-
poral implicature; still, one does not want a temporal
marker to signal a VOLITIONAL-CAUSE, even though
cause and effect are temporally related.
Finally, future work needs to address the interac-
tion of marker choice and temporal adverbs, as these
are the means to realize the simple/progressive dis-
tinction in German.
Thanks to Manfred Stede and
two anonymous reviewers for helpful comments on
earlier versions of this paper.
J. Allen. Towards a general theory of action and time.
Artificial Intelligence, 23(2), 1984.
R. B~iuerle. Temporals~itze und Bezugspunktsetzung im
Deutschen. In B. Handwerker (ed.) 1Wemde Sprache
Deutsch. Tiibingen: Gunter Narr, 1995.
J. Buscha. Lexikon deutscher Konjunktionen. Leipzig:
Verlag Enzyklop~idie, 1989.
H. Bussmaan. Lexikon der Sprachwissenschaft. Stuttgart:
KSrner, 1990.
J. Delin, D. Scott, A. Hartley. Pragmatic congruence
through language-specific mappings from semantics to
syntax. In Proe. of the 16th Conference on Computa-
tional Linguistics, Copenhagen, 1996.
B. Dorr, T. Gaasterland. Selecting tense, aspect and con-
necting words in language generation. In Proc. of the
i~th International Joint Conference on Artificial Intel-
ligence, Montreal, 1995.
V. Ehrich. The generation of tense. In G. Kempen (ed.)
Natural Language Generation: New Results in Ar-
tificial Intelligence, Psychology and Linguistics. Dor-
d.recht: Martinus Nijhoff Publishers, 1987.
M. Elhadad, K.R. McKeown. Generating connectives. In
Proc. of the 13th Conference on Computational Lin-
guistics, Helsinki, 1990.
M. Gagnon, G. Lapalme. Pr6texte: A generator for the
expression of temporal information. In Proc. of the 4th
European Workshop on Natural Language Generation,
Pisa, 1993.
R. Glas. Ein Textkorpus fiir die deutsche Gegenwart-
sprache. In: Linguistische Berichte 40, 1975, pp 63-66.
B. Grote, M. Stede. Discourse marker choice in sentence
planning. In Proc. of the 9th International Natural
Language Generation Workshop. Niagara-on-the-Lake,
Canada (to appear).
G: Helbig, J. Buscha. Deutsche Grammatik: Ein Hand-
buch fiir den Ausldnderunterricht. Berlin, Leipzig:
Langenscheidt, Verlag Enzyklop~idie, 1990.
J. Hitzeman, M. Moens, C. Grover. Algorithms for
analysing the temporal structure of discourse. In Proc.
of the Proceedings of the 6th International Conference
of the European Chapter of the Association for Com-
putational Linguistics, Dublin, 1995.
N. Hornstein. As Time Goes By. Cambridge, Mass.: MIT
Press, 1990.
A. Lascarides, J. Oberlander. Temporal connectives in a
discourse context. In Proc. of the 6th Conference of
the European Chapter of the Association for Computa-
tional Linguistics, Utrecht, 1993.
M. Moens, M. Steedman. Temporal ontology and tempo-
ral reference. Computational Linguistics, 14(2), 1988.
O. Rainbow, T. Korelsky. Applied text generation. In
Proc. of the Conference on Applied Natural Language
Processing, Trento, 1992.
H. Reichenbach. Elements of Symbolic Logic. London:
Macmillan, 1947.
D. RSsner, M. Stede. Customizing RST for the automatic
production of technical manuals. In R. Dale et al. (eds.)
Aspects of Automated Natural Language Generation.
Berlin: Springer, 1992.
M. Stede. Lexical semantics and knowledge representa-
tion in multilingual generation. Doctoral dissertation.
Published as Technical report CSRI-347, Dept. of Com-
puter Science, University of Toronto, 1996.
A. Steube. Temporale Bedeutung im Deutschen. studia
grammatica XX. Berlin: Akademie-Verlag, 1980.
K. Vander Linden, J. Martin. Expressing rhetorical rela-
tions in instructional texts: a case study of the purpose
relation. Computational Linguistics, 21(2), 1995.
L. Wanner, E. Hovy. The HealthDoc sentence planner.
In Proc. of the 8th International Workshop on Natural
Language Generation, Herstmonceux Castle, 1996.
... ¯ New lexicalizations: Most sentence planners do not produce many of the discourse-level elements found in polished texts because they are not needed when generating protosentences . For instance, discourse markers (Vander Linden and Martin, 1995; Grote, 1998) are frequently used to show the relationships between individual clauses. If appropriate information were available, it would be possible for revision operators to add discourse markers as they perform clause aggregation. ...
Full-text available
Natural Language Generation has made great strides towards multilingual gen-eration from large-scale knowledge sources. Meanwhile, current research in revision has vastly improved the qual-ity of text that NLG systems produce. However, to-date there has been no at-tempt at combining revision and mul-tilingual NLG. This paper presents re-search in multilingual revision, the last major pipelined NLG component to be studied from a multilingual perspective. We describe the linguistic difficulties in achieving multilingual revision, re-view recent work, and present an imple-mented framework for multilingual revi-sion rules.
... The statistically trained FERGUS (Chen et al., 2002) contrasts with our rule-based approach. Dorr and Gaasterland (1995) and Grote (1998) focus on generating temporal connectives, such as before, based on the relative times and durations of two events; Gagnon and Lapalme (1996) focus on temporal adverbials (e.g., when to insert a known time of day for an event). By comparison, we extend our approach to cover direct/indirect speech and the subjunctive/conditional forms, which they do not report implementing. ...
Conference Paper
We describe a method for assigning English tense and aspect in a system that realizes surface text for symbolically encoded narratives. Our testbed is an encoding interface in which propositions that are attached to a timeline must be realized from several temporal viewpoints. This involves a mapping from a semantic encoding of time to a set of tense/aspect permutations. The encoding tool realizes each permutation to give a readable, precise description of the narrative so that users can check whether they have correctly encoded actions and statives in the formal representation. Our method selects tenses and aspects for individual event intervals as well as subintervals (with multiple reference points), quoted and unquoted speech (which reassign the temporal focus), and modal events such as conditionals.
... The TimeML tool cannot be applied to Thai language which does not have tensed verb in Thai grammar. Li et al. [16] extracted the temporal relation from Chinese news by using temporal concept frames with constructed rule sets containing the explicit reference time as a temporal indicator which is the temporal marker (Grote [17] defined a temporal marker as " a word or phrase signals the temporal relation between events " ). Their temporal concept frames are linked by several events from several sentences with an explicit time expression as the time indicator. ...
Full-text available
Explanation knowledge expressed by a graph, especially in the graphical model, is essential to comprehend clearly all paths of effect events in causality for basic diagnosis. This research focuses on determining the effect boundary using a statistical based approach and patterns of effect events in the graph whether they are consequence or concurrence without temporal markers. All necessary causality events from texts for the graph construction are extracted on multiple clauses/EDUs (Elementary Discourse Units) which assist in determining effect-event patterns from written event sequences in documents. To extract the causality events from documents, it has to face the effect-boundary determination problems after applying verb pair rules (a causative verb and an effect verb) to identify the causality. Therefore, we propose Bayesian Network and Maximum entropy to determine the boundary of the effect EDUs. We also propose learning the effect-verb order pairs from the adjacent effect EDUs to solve the effect-event patterns for representing the extracted causality by the graph construction. The accuracy result of the explanation knowledge graph construction is 90% based on expert judgments whereas the average accuracy results from the effect boundary determination by Bayesian Network and Maximum entropy are 90% and 93%, respectively.
This study examined the type, frequency, and accuracy of conjunction use in a collection of Persian learners’ academic essays. To this end, the learners’ essays were graded based on the Saadi Foundation writing rubric, one of the components of which is cohesion and coherence of the text. Then a list of Persian conjunctions was extracted from four corpora: Bijan Khan, Hamshahri, Seraji, and the Saadi Foundation basic words. The conjunctions, using Fraser’s (Toward a theory of discourse markers. In: Fischer K (ed) Approaches to discourse particles. Elsevier Press, pp 189–204, 2005; Int Rev Pragm 14(2): 1–28, 2009) framework, were then classified as high-frequent, medium-frequent, and low-frequent. Afterward, 20 essays were selected from the data at each language level and the type, frequency, and correctness of the application of the conjunctions were extracted and identified. The results revealed that the use of conjunctions, both in terms of number and type, increased as language levels raised. However, no significant relationship was found between language level and the use of low-frequency conjunctions or the correct application of these conjunctions. The examination of Saadi Foundation’s writing rubric showed that the predictions made at different levels regarding discourse conjunctions were not entirely comprehensive. This study has contributed to the field by developing a table of the most widely used conjunctions in each level of Persian proficiency.
Narrative generation has historically suffered from poor writing quality, stemming from a narrow focus on story grammars and plot design. Moreover, to-date natural language generation systems have not been capable of faithfully reproducing either the variety or complexity of naturally occurring narratives. In this article we first propose a model of narrative derived from work in narratology and grounded in observed linguistic phenomena. Next we describe the Author architecture for narrative generation and an end-to-end implementation of the Author model in the StoryBook narrative prose generation system. Finally, we present a formal evaluation of the narratives that StoryBook produces.
Conference Paper
Full-text available
We show that discourse structure need not bear the full burden of conveying discourse relations by showing that many of them can be explained nonstructurally in terms of the grounding of anaphoric presuppositions (Van der Sandt, 1992). This simplifies discourse structure, while still allowing the realisation of a full range of discourse relations. This is achieved using the same semantic machinery used in deriving clause-level semantics.
Conference Paper
Full-text available
Pipelined Natural Language Generation (NLG) systems have grown increasingly complex as architectural modules were added to support language functionalities such as referring expressions, lexical choice, and revision. This has given rise to discussions about the relative placement of these new modules in the overall architecture. Recent work on another aspect of multi-paragraph text, discourse markers, indicates it is time to consider where a discourse marker insertion algorithm fits in. We present examples which suggest that in a pipelined NLG architecture, the best approach is to strongly tie it to a revision component. Finally, we evaluate the approach in a working multi-page system.
Full-text available
We show that discourse structure need not bear the full burden of conveying discourse relations by showing that many of them can be explained nonstructurally in terms of the grounding of anaphoric presuppositions (Van der Sandt, 1992). This simplifies discourse structure, while still allowing the realisation of a full range of discourse relations. This is achieved using the same semantic machinery used in deriving clause-level semantics. 1 Introduction Research on discourse structure has, by and large, attempted to associate all meaningful relations between propositions with structural connections between discourse clauses (syntactic clauses or structures composed of them). Recognising that this could mean multiple structural connections between clauses, Rhetorical Structure Theory (Mann and Thompson, 1988) simply stipulates that only a single relation may hold. Moore and Pollack (1992) argue that both informational (semantic) and intentional relations can hold between clauses simul...
Full-text available
this paper on interactions among these elements, since we feel that there is still much that needs to be learned about them from empirical and experimental studies. However, our use of anaphoric presupposition in an account of discourse relations suggests a unified account of discourse connectives, tense (whose anaphoric nature has long been argued for -- cf. Partee (1984) and Webber (1988) inter alia), modality (which Stone (1999) proposes to also treat anaphorically, parallel to tense), presuppositional determiners such as "other", "another", "similar", etc., and focus particles such as "even" and "only", which Stede has suggested can sometimes be used to convey the same meaning as a discourse connective, as in
Full-text available
In text, temporal relations between events can be signalled in several ways; among them are specific lexical items, here called temporal discourse markers. We analyse the semantics of about 20 German subordinating conjunctions and prepositions and transfer these findings to a sentence generation framework that uses a dedicated discourse marker lexicon for producing complex sentences. After discussing the ontological decisions and the lexical representations, we demonstrate how this information can be used to choose an appropriate temporal marker when verbalizing a pair of time-stamped event representations. 1 Introduction In knowledge-based natural language generation (NLG), an abstract meaning representation is successively transformed into a linguistic sentence or text. Specifying this mapping involves both designing the abstract representation and defining its relationship with lexical knowledge, which involves ontological decisions. In this paper, we investigate the specific pro...
Conference Paper
Full-text available
We present a study of the mappings from semantic content to syntactic expression with the aim of isolating the precise locus and role of pragmatic information in the generation process. From a corpus of English, French, and Portuguese instructions for consumer products, we demonstrate the range of expressions of two semantic relations, GENERATION and ENABLEMENT (Goldman, 1970) in each language, and show how the available choices are constrained syntactically, semantically, and pragmatically. The study reveals how multilingual NLG can be informed by language-specific principles for syntactic choice.
Conference Paper
Full-text available
In this paper we present a method for generating French texts conveying temporal information. This method integrates the Discourse Representation Theory (DRT) and the Systemic Grammar Theory. First, we show how the DRT is used to represent temporal information. We then show how this formalism is used to represent the temporal localization expressed by temporal adverbial phrases. Finally, we give a description of how this representation of adverbial phrases can be translated into a syntactic form, using Systemic Grammar Theory. Prétexte, our implementation of this method, is able to generate a great variety of temporal adverbial phrases.
Full-text available
This paper addresses this issue in the context of the expression of procedural relations between actions in instructional text. It employs the following four step approach to achieve this goal: (1) Collect a corpus of the relevant text type; (2) Perform a detailed linguistic study of a portion of this corpus, called the training set, and reserving the remainder as a testing set; (3) Implement the results of this study in a text generation system; (4) Compare the output of the system with the text found in the entire corpus. This has resulted in the construction of IMAGENE, an instructional text generation system which embodies a model of the forms of expression consistently used by instructional text writers over a broad range of instruction types. The details of IMAGENE's
In this chapter I am concerned with the question of how speakers convey temporal meaning in actual speech. More specifically, I will be dealing with speakers’ choice of tense forms contributing appropriately to the linguistic formulation of the intended message. Although tense marking of the verb is obligatory in all Indo-European languages, tense is not the only linguistic means for conveying temporal information in these languages, nor is the meaning of the tenses the only type of temporal information contributing to the temporal reference of an utterance. Therefore, I will first present an overview of different types or categories of temporal meaning and relate these to certain categories of linguistic expressions (Section 1). In Section 2, I will present some observations about the interaction between tense, lexical aspect and temporal adverbs in German. Finally, in Section 3, I will discuss some of the constraints that these interactions impose on a language production system which is functioning at well-defined levels of processing (conceptual, functional, positional) and is characterized by specific processing properties like modularity, incrementality and linearity.
Conference Paper
Rhetorical Structure Theory (RST) has emerged as a promising candidate for text representation in NLG. We investigated the applicability of RST in the automatic production of multilingual technical manuals. Starting from a domain knowledge base, we construct an RST-tree for a particular manual section, which is then converted to a set of sentence plans. These plans serve as input to sentence generators that produce the final text. In this paper, we report first on a number of open questions regarding general aspects of RST. Arguing that the original set of RST relations is not specific enough for practical generation purposes, we suggest a number of new relations that we found useful in our domain. After briefly examining the stage of RST tree construction, we then outline a procedure for converting RST trees to a sequence of sentence plans.
A formalism for reasoning about actions is proposed that is based on a temporal logic. It allows a much wider range of actions to be described than with previous approaches such as the situation calculus. This formalism is then used to characterize the different types of events, processes, actions, and properties that can be described in simple English sentences. In addressing this problem, we consider actions that involve non-activity as well as actions that can only be defined in terms of the beliefs and intentions of the actors. Finally, a framework for planning in a dynamic world with external events and multiple agents is suggested.