Content uploaded by Cristina Grisot
All content in this area was uploaded by Cristina Grisot on Oct 25, 2020
Content may be subject to copyright.
Temporal Coherence in Discourse: Theory
and Application for Machine Translation
Abstract Temporal coherence in discourse is provided through several temporal
cohesive ties, such as tense, aspect and discourse connectives. In the relevance
theoretic framework and more speciﬁcally in the Geneva school of pragmatics,
these cohesive ties are considered as encoding procedural information important for
guiding the hearer towards the intended interpretation of the discourse. Jacques
Moeschler and his team studied temporal cohesive ties and proposed original
theoretical models that have been validated with human and automatic annotation
experiments, as well as in language acquisition studies (Zufferey and
Popescu-Belis, this volume). In this paper, I show that Jacques Moeschler’s model
for inferring temporal discourse relations and his description of tenses expressing
past time in French is cross-linguistically valid and can be modelled for improving
the results of statistical machine translation systems.
Keywords Verbal tenses ⋅Discourse coherence ⋅Natural language processing ⋅
Machine translation ⋅Cross-linguistic ⋅Relevance theory ⋅Cohesive ties
The ideas presented in this paper are based on a series of articles published with Bruno Cartoni,
Thomas Meyer, Andrei Popescu-Belis, Michele Costagliola and Jacques Moeschler, with whom
I collaborated on two research projects. I am very grateful to Jacques Moeschler for his guidance
and resourceful discussions since the beginning of my research. A theoretical model of temporal
reference in tensed languages based on empirical work (corpus-based and experiments with
native speakers) is proposed in my Ph.D. dissertation (Grisot 2015), in which I make the
proposal that temporal coherence in discourse is triggered by the hearer’s need to acquire
temporal coherence at the cognitive level. In order to attain this purpose, he treats information
coming from several sources (tense, grammatical aspect, lexical aspect, temporal adverbials and
temporal connectives) and their rich interrelations in a coherent manner.
C. Grisot (✉)
Département de Linguistique, Université de Genève, Geneva, Switzerland
© Springer International Publishing AG 2017
J. Blochowiak et al. (eds.), Formal Models in the Study of Language,
Language Processing (NLP), Machine Translation (MT), Semantics and
Pragmatics are ﬁelds that have had an increasing interest in linguistic phenomena
such as discourse connectives, pronoun anaphora and verb tenses due to their
important role for discourse coherence. The essential feature that makes a piece of a
text a discourse is the coherent succession of sentences forming a whole and
referring to the same entities (nominal or eventualities). However, it seems that this
condition is not sufﬁcient, as shown in example (1) (Hobbs 1979:67), which is not
coherent even if “he”can refer to “John”. Hobbs argues that there is an expectation
of coherence, which is deeper than the notion of a discourse just being “about”
some set of entities. Sentences in (2) and (3) are coherent because the hearer infers a
causal relation in (2) and a temporal relation in (3).
(1) John took a train from Paris to Istanbul. He likes spinach.
(2) John took a train from Paris to Istanbul. He hates airplanes.
(3) John took a train from Paris to Istanbul. He went by boat from there to Cyprus.
The Geneva School of Linguistics,
and Jacques Moeschler in particular,
investigated linguistic expressions on which discourse coherence depends, such as
French connectives and verb tenses. Zufferey and Popescu-Belis (this volume)
discuss how Moeschler’s classiﬁcation and description of the role of French dis-
course connectives and of discourse relations have been validated empirically in
language acquisition studies, NLP and MT.
In this paper, I will discuss Moeschler’s proposal for inferring temporal and
causal discourse relations and show that his model is cross-linguistically valid and it
can be used for improving the results in terms of coherence of statistical machine
translation systems (SMT). I will concentrate particularly on verb tenses in English
(EN), French (FR), Italian (IT) and Romanian (RO) and their usage in multilingual
translation corpora. I review theoretical accounts of temporal coherence in dis-
course (classical and pragmatic descriptions) in Sect. 2. Section 3is dedicated to
the link between verb tenses expressing past time and temporal coherence in dis-
course and to pragmatic features proposed for empirical testing. Section 4provides
The COMTIS Project (Improving the Coherence of Machine Translation Output by Modeling
Intersentential Relations; project n° CRSI22_127510, March 2010–July 2013) and the MODERN
Project (Modeling discourse entities and relations for coherent machine translation; project n°
CRSII2_147653, August 2013–August 2016) belong to the Sinergia interdisciplinary program
funded by the Swiss National Science Foundation.
At the beginning of the eighties, the label “Geneva School”was given to a series of publications
on discourse and conversation that applied basic principles of syntactic analysis to the domain of
discourse (Roulet et al. 1985; Moeschler 1985). In the beginning of the nineties, two different
directions could be identiﬁed in the Geneva School: (i) a general discourse-oriented framework of
language based on the modular hypothesis (Roulet 1997) and (ii) a radical pragmatic perspective
on discourse sequencing and discourse interpretation (Moeschler 1993,1996) (see detailed pre-
sentation in Moeschler 2001).
356 C. Grisot
the results of corpus analysis, human and automatic annotation of data, as well as
the improvement in the translation of EN Simple Past (SP) into FR by automatic
systems. Finally, Sect. 5concludes this paper.
2 Temporal Coherence in Discourse: Theoretical
2.1 Discourse Relations
In theoretical linguistics, studies have aimed at describing the factors that contribute
to discourse coherence and categorizing the different types of coherence relations
that connect clauses and sentences. Halliday and Hasan (1976) proposed the terms
cohesive ties and cohesion for the linguistic devices used to build coherence
between sentences. A number of theories made use of relations in explaining
coherence (e.g. Longacre 1983; Hobbs 1985; Grosz and Sidner 1986; Mann and
Thompson 1988). Relations that link clauses are known as “rhetorical predicates”
(Grimes 1975), “conjunctive relations”(Halliday and Hassan, 1976)or“intentions”
(Grosz and Sidner 1986). The term of “coherence relations”is due to Hobbs (1979).
From the theoretical linguistics perspective, the taxonomy of discourse relations
consists of several types, the most well known are those proposed by Halliday and
Hassan: additive, temporal, causal and adversative (contrast). Kehler (2004: 244)
points out that “an explanatory theory of coherence requires a set of externally
driven principles to motivate and ultimately constrain the relation set.”This is the
direction taken by Sanders and colleagues (Sanders et al. 1992; Sanders and
Noordman 2000; Sanders 1997,2005) by proposing a theory in which psycho-
logical plausibility is the primary motivating factor. In their view, discourse rela-
tions are composed of more ﬁne-grained and more primitive features than in the
previous taxonomies, such as basic operation (causal and additive meanings), order
of segments (basic or non-basic), polarity (positive or negative) and source of
coherence (semantic and pragmatic). Thus, Sanders et al. proposed a principled and
explanatory theory of coherence, which leaves open the possibility to add other
factors that interact with the listed ones.
Following Hobbs (1979), Sanders and colleagues (1992) state that coherence
relations account for coherence in the cognitive representation of a discourse and
they see coherence relations as cognitive entities (Mann and Thompson 1986;
Sanders et al. 1992). Sanders (2005) suggested that when language users process a
discourse, they connect discourse segments by inferring coherence relation on the
basis of a very limited set of cognitive principles, such as causality and subjectivity.
At a more general level, he aimed at shedding light on human cognition by
investigating the mechanisms underlying discourse coherence. According to
Sanders, the causality and subjectivity cognitive principles account for the use of
linguistic expressions (connectives and other lexical items), discourse coherence
Temporal Coherence in Discourse: Theory …357
through discourse relations, language acquisition and discourse processing. Sanders
et al. (1992) argued that the essential property of cognitive discourse relations is
that they establish coherence in the cognitive representation language users have or
make of a discourse.
In a different framework, pragmatic theories (Grice 1989, neo
trends, Geneva School) adopted the idea that human communication in general, and
discourse coherence in particular, are inferential processes driven by the desire to
express and recognize intentions. In this paper, I am interested in temporal
coherence in discourse and its cohesive ties. A question that was asked in the
literature is how are temporal relations in discourse inferred. Answers have been
proposed by several trends: Discourse Representation Theory (DRT: Kamp and
Reyle 1993) and Segmented Discourse Representation Theory (SDRT: Asher 1993;
Lascarides and Asher 1993; Asher and Lascarides 2003), Relevance Theory (RT:
Wilson and Sperber 1998) and Geneva School (Reboul and Moeschler 1998;
Moeschler 2000,2002;2005). I will discuss these approaches in what follows.
Since SDRT, it is generally accepted that discourse relations are inferred on the
basis of discourse,linguistic and world types of knowledge. Let’s consider two
temporal discourse relations: Narration and Explanation (Asher 1993). The Nar-
ration discourse relation is characterized by a forward temporal inference, i.e. time
advances while Explanation is characterized by a backward temporal inference, i.e.
time goes backward. In what follows, I will brieﬂy describe these types of
knowledge as they are approached in discourse semantics.
Firstly, discourse knowledge is provided by discourse type, for example, nar-
rative or non-narrative discourses. Labov and Waletztky (1967) argued that Nar-
ration is highly preferred in narrative discourses whereas Explanation is preferred in
Secondly, linguistic knowledge is provided by temporal connectives and verb
tenses. Kamp and Rohrer (1983) argued that French verb tenses expressing past
time encode time direction necessary for inferring discourse relations. For example,
the Passé Simple (PS) encodes a forward temporal inference, the Plus-que-parfait
encodes a backward inference and the Imparfait (IMP) encodes an inclusive tem-
poral inference. Unfortunately, this idea has numerous counterexamples discussed
by Kamp and Rohrer (1983) themselves as in (4), Moeschler (2000), Saussure
(1997,2000; Tahara 2000 for the PS; Saussure and Sthioul 1999,2005 for the IMP)
(4) Bianca chanta le récitatif et Igor l’accompagna au piano.
‘Bianca sung the recitative and Igor accompanied her on the piano.’
For example Gazdar (1979), Horn (1984;1992;2004;2007), Levinson (1983;2000).
For example Sperber and Wilson (1986/1995), Blakemore (1987,2002), Carston (2002).
358 C. Grisot
Thirdly, world knowledge is stronger and has priority over linguistic knowledge
for determining the inferred temporal discourse relation, as in (5). According to
Kamp and Rohrer, the PS encodes a forward inference, which should hold in both
examples (a) and (b). But in (b), there is a backward inference based on the causal
relation that comes from world knowledge.
(5) (a) Max poussa Jean. Il tomba.
‘Max pushed John. He fell’
(b) Jean tomba. Max le poussa.
‘John fell. Max pushed him.’
The weakness of this approach is circularity: discourse type is deﬁned based on
the usage of the appropriate verb tenses corresponding to the intended temporal
inference (forward or backward) and discourse relation is inferred based on the
discourse type. Moeschler (2005) points out that the most appropriate explanation
for establishing temporal inferences in discourse is a pragmatic one. Speciﬁcally,
we need models that explain how linguistic and contextual information are com-
bined. The following section is dedicated to pragmatic models of temporal
2.2 Pragmatic Models
One of the ﬁrst pragmatic propositions for explaining temporal inferences in dis-
course is based on one of Grice’s(1967) maxims “Be orderly”. In the gricean
framework, temporal inferences are conversational implicatures triggered by the
respect or lack of respect of conversational maxims. In other words, these infer-
ences correspond to beliefs attributed to the speaker who is expected to have obeyed
the Cooperative Principles and the maxim of order. Gricean conversational impli-
catures have no contribution to the truth conditions of utterances. Cohen (1971) was
the ﬁrst to question the treatment of temporal inferences as conversational impli-
catures by showing their link to connectives such as and, because and but. Fur-
thermore, relevance theorists (Sperber and Wilson 1986/1995; Carston 1988,1993;
Wilson and Sperber 1998) treat temporal inferences as “pragmatically determined
aspects of what is said”(Wilson and Sperber 1998: 172). In other words, temporal
inferences are part of the explicature of an utterance, and they provide the same
status to causal relations. Moeschler (2000) argued that causality plays an important
role for temporal coherence in discourse and he proposed a theoretical model for
temporality that includes this parameter. As it will be shown in Sect. 4.2,
Moeschler’s theoretical model was validated experimentally in ofﬂine experiments
with linguistic judgement tasks.
Moeschler proposed several arguments in favour of treating temporal relations as
part of explicatures. I will brieﬂy recall them below.
Temporal Coherence in Discourse: Theory …359
Firstly, the temporal interpretation corresponds to a pragmatic enrichment of the
propositional form of the sentence and they contribute to its truth conditions. In
example (6) from Wilson and Sperber (1998: 171), the disjunction is not redundant
because each disjunct brings a genuine contribution to the truth-conditions of the
utterance. This is based on the assumption that the events presented in each disjunct
happened in a different order.
(6) It’s always the same at parties: either I get drunk and no-one will talk to me or
no-one will talk to me and I get drunk.
Secondly, relevance theorists’explanation focuses on processing efforts rather
than on cognitive effects. Example (5)(a) produces two interpretations (either for-
ward temporal inference or backward causal inference) and neither syntactic nor
semantic structures indicate how the sentence should be interpreted. Moeschler
argued that the interpretation is consistent with the cognitive relevance principle.
This means that a temporal or a causal interpretation will be chosen depending on
which manifest facts are more accessible to the hearer and based on the mutual
Thirdly, forward temporal inference (called temporal sequencing) and backward
causal inference (called reverse-causal interpretation) are not the only possible
relations among eventualities. There are two other possible relations, namely si-
multaneity as in (7) and indeterminacy as in (8).
(7) Bill smiled. He smiled sadly.
‘Bill souriait. Il souriait tristement.’
(8) That night, our hero consumed half a bottle of whisky and wrote a letter to Lady
‘Cette nuit-là, notre héros but la moitié d’une bouteille de whisky et écrivit une
lettre à Lady Ann.’
The relations of simultaneity and indeterminacy are deﬁned as follows:
(A) E1 covers (partially) e2 is a part of the eventuality denoted by e1 is included in
the temporal interval deﬁning e2
(B) The relation between e1 and e2 is undetermined if the determining the relation
is not necessary for understating e1 and e2 or if determining the relation is not
Fourthly, temporal sequencing (i.e. forward temporal inferences) does not seem
to be central for temporal coherence in discourse. Causality plays an important role
therefore the question concerning the relation between temporality and causality
should be asked. In example (9), the only possible relations are forward causal and
forward temporal relations whereas in (10) several relations are possible: forward
temporal and causal, forward temporal and backward causal, backward temporal
and causal. Wilson and Sperber (1998) give an example where a causal relation
occurs without a temporal relation as in (11).
360 C. Grisot
(9) Socrate but un coup et tomba raide.
‘Socrate drank and fell stone.’
(10) Marie cria et Pierre partit.
‘Mary screamed and Peter left.’
(11) Susan is underage and can’t drink.
Hence, Moeschler (2000) proposed that causal and temporal relations are two
sets of relations that can have a Boolean junction. This means that for two even-
tualities e1 and e2, there can exist an intersection of causal and temporal relations
for which [e1 causes e2] implicates [e1 precedes e2]. He also suggested that two
sentences can produce identical cognitive effects on the basis of different explica-
tures and implicated premises as in (12) and (13). In (12), the temporal relation [e1
precedes e2] is part of the explicature while the causal relation [e1 causes e2] is an
implicated premise. In (13), the causal relation [e1 causes e2] is part of the
explicature while the temporal relation [e1 precedes e2] is part of the implicated
(12) Max a laissé tomber le verre (e1). Il s’est cassé (e2).
‘Max dropped the glass. It broke.’
(13) Le verre s’est cassé (e2). Max l’a laissé tomber (e1).
‘The glass broke. Max dropped it.’
It was stated that there are several possible types of relations among eventual-
ities, and this can be summarized in Fig. 1from Moeschler (2000). The model
considers temporal and causal relations. As far as temporal relations are concerned,
they can be or not forward temporal inferences (temporal sequencing). In the case
Fig. 1 Possible relations among eventualities
Temporal Coherence in Discourse: Theory …361
where there is no temporal sequencing, there are two new possibilities: either there
is or is not a backward temporal inference. And ﬁnally, if there is no backward
temporal inference, then the cases of temporal simultaneity or indetermination can
be identiﬁed. Temporal sequencing can be accompanied or not by a forward causal
relation, as in (9) and (11) respectively. Backward temporal inference can be
accompanied or not by reverse causality, as in (14) and (15) respectively.
(14) Max tomba. Jean l’avait poussé.
‘Max fell. John had pushed him.’
(15) Jean prepara son café. Il s’était levé sans entrain.
‘Jean prepared his coffee. He woke up without energy.’
Moeschler (2000) concludes that the natural way of presenting eventualities is
not as a backward temporal inference, unless it is speciﬁed otherwise through the
verb tense used and discourse connectives for example. He suggests that temporal
or causal interpretations of a sentence are triggered by the speaker’s intention. The
speaker may have a communicative intention of presenting the eventualities as they
occurred and thus the interpretation consistent with the cognitive principle of rel-
evance is the temporal one. The speaker may have a communicative intention of
presenting eventualities focusing on the causal relation holding among them. In this
case, the interpretation consistent with the principle of relevance is the causal one.
3 Temporal Coherence and Verb Tense
A contrastive analysis of verb tenses used in the examples given in Sect. 2.2 shows
that FR is much more sensitive than EN to the type of relation expressed. If the
EN SP can be equally used for forward and backward temporal and causal relations
as well as temporal simultaneity, FR tenses seem more specialized. The FR PS
seems specialized for the forward temporal relation (example (9)), the
Plus-que-Parfait in backward temporal relations (examples (14) and (15)) and the
IMP in the case of temporal simultaneity (example (7)).
The Passé Composé
(PC) does not impose any constraint relating temporal and causal inferences (ex-
amples (12) and (13)).
At this point of the discussion, I would like to introduce the conceptual/procedural
distinction referring to types of information encoded by linguistic expressions. The
conceptual/procedural distinction was proposed by Blakemore (1987) to explain
differences between words with a conceptual content, such as table, cat, think or walk
on the one hand, and discourse connectives, such as but, so or also. Content words
I assume that this specialization conducted Kamp and Rohrer (1983) to argue that the PS encodes
a forward temporal inference. I argue that the PS encodes a procedure regarding directional
temporal inference. In other words, it is an instruction for the hearer to ascertain the contextual
value of the directional temporal inference.
362 C. Grisot
encode concepts that contribute to the proposition expressed by an utterance while the
meaning of a discourse connective is better described in terms of constraints on the
inferential phase of interpretation than in conceptual terms. The hearer is expected to
have access to the smallest and most accessible set of contextual assumptions in order
to get the intended cognitive effects.
Verb tenses also have been investigated regarding the conceptual/procedural
distinction and their role for discourse processing (Wilson and Sperber 1998;
Moeschler 1993; Ahern and Leonetti 2004 for the Spanish subjective; Nicolle 1997,
1998; Moeschler et al. 1998;2012; Saussure 2003; Leonetti and Escandel-Vidal
2003 on the Spanish imperfective; Aménos-Pons 2010,2011 on Spanish past
tenses, to name but a few). Similarly to connectives, verb tenses are considered to
encode procedural information consisting of instructions on how to manipulate
mental representations of eventualities. Speciﬁcally, they give information about
the temporal and causal relations holding among eventualities.
Since the ﬁrst studies on verb tenses in a relevance theoretic framework, it is
believed that they have rigid procedural meanings that help the hearer reconstruct
the intended representation of eventualities (Nicolle 1998; Aménos-Pons 2011;
Saussure 2003,2011). Saussure (2003) proposes algorithms to follow, consisting of
the instructions encoded by verb tenses, in order to grasp the intended meaning of a
verb tense at the discourse level. Taking the distinction conceptual-procedural as a
foundation, Blakemore (1987), Wilson and Sperber (1993), Moeschler (1994,
1998) and Nicolle (1997,1998) claimed that tenses have a procedural meaning.
Nicolle (1998: 4) argues that tense markers impose constraints on the determination
of temporal reference and thus they “may be characterized as exponents of pro-
cedural encoding, constraining the inferential processing of conceptual represen-
tations of situations and events”. Concerning the status of the temporal coordinates,
Saussure and Morency (2012) argue that tenses encode instructions on how the
eventuality is to be represented by the hearer through the positions of temporal
coordinates. They consider thus that temporal location with the help of point of
speech S, reference moment R and event moment E (Reichenbach 1947)isofa
A different view was defended in Grisot and Moeschler (2014). We argued that
that location through temporal coordinates S and E does not constrain the inferential
processing but contributes to the propositional content of the utterance. Seen from
this perspective, temporal coordinates are conceptual parameters saturated con-
textually. More speciﬁcally, the hearer is brought to build an ad hoc concept as
pastness or non-pastness based on contextual information.
We considered Wilson
and Sperber’s example (1993: 157) given in (16) and the propositional form given
in (17). We add to this propositional form the information that eventualities of
saying and of being tired took place before the moment when the sentence was
For a development of the conceptualist view of tense based on experimental ﬁndings using the
cognitive features of conceptual and procedural information proposed by Wilson and Sperber
(1993), see Grisot (2015).
Temporal Coherence in Discourse: Theory …363
uttered. The extended propositional form would be something like the one given in
(18). This temporal information cannot be cancelled or contradicted, as shown by
the incompatibility with the adverb now or tomorrow in (19) and (20), as well as the
compatibility with the adverb yesterday in (21).
(16) Peter told Mary that he was tired.
(17) Peter Brown told Mary Green at 3.00 pm on June 23 1992 that Peter Brown
was tired at 3.00 pm on June 23 1992.
(18) Peter Brown told Mary Green at 3.00 pm on June 23 1992 (a moment before
the present moment/in the past) that Peter Brown was tired at 3.00 pm on June
23 1992 (a moment before the present moment/in the past).
(19) *Peter Brown told Mary Green at 3.00 pm on June 23 1992 which is now (a
moment contemporary with the moment of speech)/tomorrow (a moment
which is after the moment of speech) that Peter Brown was tired at 3.00 pm on
June 23 1992 which is now/tomorrow.
(20) *Now/tomorrow Peter told Mary that he was tired.
(21) Yesterday, Peter told Mary that he was tired.
In what follows, I will focus on the procedural content of verb tenses and their
link to temporal coherence in discourse. Procedural information consists of
instructions and constraints for contextual usages of a tense leading to the inter-
pretation intended by the speaker. As noted in Sect. 2.2, French verb tenses
expressing past time seem specialized in expressing forward and backward tem-
poral and causal relations. Moeschler (2002) makes the hypothesis that this “spe-
cialization”is due to the procedural content encoded by French verb tenses.
Speciﬁcally, one type of procedural information concerns temporal and causal
relations among eventualities. In Grisot and Moeschler (2014), this procedure is
called the [± narrative] feature.
is a binary pragmatic feature: in
narrative usages, a verb tense expresses eventualities (events/states
) that are tem-
porally ordered accompanied or not by a causal relation, while non-narrative usages
express temporal simultaneity or temporally un-related states of affairs.
The narrativity feature is a coarse grained semantic and pragmatic feature proposed in a particular
framework, that of Natural Language Processing and Machine Translation. I admit that from a
theoretical point of view, only ﬁner coarse features can explain all possible (both very frequent and
less frequent) usages of a verb tense.
Verb tense has frequently been associated with narrative contexts in various frameworks, such as
in DRT and SDRT. Smith (2003) discussed discourse modes based in textual structure and aspect.
If these theories focused on linguistic information and made use of non-monotonic inferences, for
us narrativity is procedural information representing a cognitive (as opposed to logic in (S)DRT)
discourse relation (Hobbs 1979; Mann and Thompson 1988; Sanders et al. 1992). Cognitive
discourse relations are expressed lexically through verb tense and connectives (that are
language-speciﬁc) and can occur in any type of stylistic register.
At this point of the research, I consider lexical aspect as one class and do not distinguish between
Vendler (1957). Moeschler (2000) discusses Dowty’s principle of interpretation of temporal dis-
courses (1986) based on lexical aspect. Moeschler (2000) argues that this approach to temporal
relations adopts a radical position and does not explain a certain number of exceptions.
364 C. Grisot
If we go back to Fig. 1presenting the possible relations between two eventu-
alities e1 and e2, we can observe that a narrative usage of a tense assembles
temporal sequencing (with forward or no causal relation) and backward temporal
sequencing (with reverse or no reverse causality). A non-narrative usage of a tense
corresponds to temporal simultaneity. As far as the last possible relation is con-
cerned, Moeschler (2000) suggests that two cases are possible. Firstly, one of the
three previous relations can be determined on the basis on contextual assumptions
and mutual cognitive environment. Secondly, no relation can be determined. The
case of un-related eventualities was included under non-narrative usages of a tense.
Our research in the COMTIS and MODERN projects aimed at identifying
semantic and pragmatic features that would improve the translation of verb tenses
by statistical machine translation systems. In order to achieve this purpose,
empirical research was carried out consisting of corpus work (Sect. 4.1), followed
by human and automatic annotation experiments with the [± narrative] feature
4 Empirical Work
4.1 Corpus Work and Verb Tense Translation Divergences
Grisot and Cartoni (2012) studied the discrepancies between theoretical descrip-
tions of verb tenses and their use in parallel corpora. The corpus consists of texts in
EN and their translations into FR belonging to four different genres with the fol-
lowing distribution: literature (18 %), journalistic (18 %), parliamentary discussions
(31 %) and legislation (33 %). A total of 1275 predicative verb tenses have been
considered, which represents 77 % of the verb tenses occurring in the corpus.
Corpus analysis was done in two steps: (i) a ﬁrst monolingual step in order to
calculate the frequency of verb tenses in a source language (SL), and (ii) a second
bilingual step in order to identify the tenses used as translation possibilities into a
target language (TL) for a certain tense in a SL. Calculating the frequency of tenses
in the corpus allowed us to verify if verb tenses that are considered to be
ambiguous, are also frequent in corpora. Quantitative analyses of tokens of SP in
our corpora enabled us formulate statistically signiﬁcant observations.
The monolingual analysis of the corpus containing texts in EN as SL showed
that the most frequent tenses are: the Simple Present (32 %), the SP (25 %) and the
Present Perfect (14 %). The contrastive bilingual analysis of the parallel corpus
revealed that the SP is one of the most ambiguous EN verb tenses, as far as its
translation into FR is considered. The FR tenses used to render the semantic and
pragmatic meaning of the SP are: the IMP used most often in literature (44 %), the
PC used most frequently in the journalistic register (58 %), the PS used most
Temporal Coherence in Discourse: Theory …365
frequently in the literature register (40 %). The Present tense is used in the legis-
lation register (10 %) in order to create a deontic effect.
This distribution shows
that genre is not a good predictor, as it could have been expected, i.e. in the
literature genre the SP is translated with an IMP in 44 % of the cases and with a PS
in 40 %.
In another corpus-based study (Grisot and Costagliola, 2014), we studied the
translation of the EN SP into Italian (IT) and Romanian (RO). The parallel corpus
consists of texts from four different genres with the following distribution: literature
(37 %), journalistic (18 %), parliamentary discussions (19 %) and legislation
(26 %). There are three most frequent verbal tenses used in these Romance lan-
guages for the translation of the SP, as shown in Table 1.
A question that could be asked is how can the translation divergence of the SP be
explained? Two answers are possible:
(A) The SP is polysemantic –which means it has several meanings and each
meaning is translated into FR, IT or RO through a different tense.
(B) The SP is underdetermined –which means that the meaning of the SP must be
contextually worked out by assigning contextual values to both conceptual and
procedural information. The consequence of the underdetermined meaning of
the SP is that it has several contextual usages and each contextual usage is
translated into FR, IT or RO through a different tense.
As suggested by Smith (1990) and based on the results of the empirical work, we
adopted the second possibility. To account for the translation divergence of the SP,
Grisot and Moeschler (2014) made the hypothesis that the procedural feature [±
narrative] can be used for disambiguating among different possible usages of the
SP. As mentioned in Sect. 2, the link between verb tenses and temporal coherence
in discourse has been discussed for FR verb tenses. One of the advantages of
translation corpora is to permit cross-linguistic transfer
of semantic and/or
Table 1 Translation
possibilities of the SP into FR,
IT and RO in the corpus
FR (%) IT (%) RO (%)
PC 37 33 49
IMP 24 18 15
PS 16 22 18
Others 16 21 13
The translation of the SP through a present tense form can be explained by the contextual values
taken by temporal coordinates S, R and E in order to lead to the speaker’s intended interpretation.
Speciﬁcally, the translation with present time signals that the eventuality is viewed from the
moment of speech (R = S) (see Grisot and Moeschler, submitted for publication, where we argue
based on experimental results that temporal coordinates are conceptual information).
Samardzic (2013) uses this novel methodology for investigating the translation equivalents of a
range of English light verb constructions into several languages. Slavic languages encode verb
aspect lexically, unlike other European languages. She applies the aspectual representation
obtained in the English-Serbian cross-linguistic setting to classify English verbs into event
366 C. Grisot
pragmatic information. This is due to the fact that studying instances and usages of
verb tenses in a parallel corpus make it possible to control for context and register
In the two experiments we designed, we made use of the methodology of
cross-linguistic transfer of properties. Speciﬁcally, the [±narrativity] feature was
tested and validated in two experiments with human judges. The working
hypotheses of the ﬁrst experiment were formulated based on theoretical descriptions
of FR tenses. The second experiment was designed based on the cross-linguistic
transfer of semantic and pragmatic information from FR past tenses to the
EN SP. These experiments resulted in the creation of human annotated data that
served as training data for an automatic classiﬁer. The automatic classiﬁer at its turn
was integrated into a statistical machine translation system.
This interdisciplinary research from the COMTIS and MODERN projects lead to
a series of publications, as follows. The human experiments are presented in detail
in Grisot and Moeschler (2014). The NLP work described in Grisot and Meyer
(2014) consists of the automatic labelling of the corpus with the narrativity feature.
And ﬁnally, the MT work and its results are described in Meyer et al. (2013). In the
following section, I will give a brief summary of the experiments, both for human
and automatic annotation, as well as the implementation of the model for improving
the results of statistical machine translation systems. I argue that the ﬁne-grained
theoretical model proposed by Moeschler (2000,2002) for temporal coherence is
empirically and cross-linguistically valid.
4.2 Human Judgement, Application to NLP and Improving
the Results of MT
Kamp and Rohrer (1983) argued that temporal sequencing is semantic information
encoded by verb tenses: the PS makes time to advance whereas with the IMP time
does not advance (i.e. it is a verbal tense of background information). In other
words, the PS has a narrative usage and the IMP has a non-narrative usage.
Experiment 1 was designed to test this theoretical assumption. Seventy-six FR
native speakers participated in this experiment and were instructed to judge sen-
tences randomly chosen (15 sentences per participant) from the FR part of the
corpus regarding the [±narrative] feature. The two possible values of feature were
explained and illustrated with several examples. The sentences contained one of the
three FR tenses of interested: PC, PS and IMP. The results
of the experiment
These results include only cases where inter-annotator agreement is high. Four participants
judged each sentence and the result is based on the majority of answers.
Temporal Coherence in Discourse: Theory …367
(A) The PS was judged as having a narrative usage in 92 % of the cases;
(B) The PC was judged as having a narrative usage in 77 %;
(C) The IMP as having a non-narrative usage in 77.5 %.
This leads to about 23 % of the cases where the PC has a non-narrative usages
and the IMP has a narrative usage. These results are consistent with the theoretical
descriptions (Moeschler 2000,2002; Saussure and Sthioul 1999,2005 for the IMP;
Saussure 1997 for the PS; Lusher and Sthioul for the PC). Moeschler (2002)
suggests that the PS is specialized for forward temporal and causal and it imposes
this interpretation even in cases when lexical information (order of eventualities) is
not compatible. The PC is not specialized for directional temporal and causal
relations allowing all possible relations presented in Fig. 1. The IMP is expected to
occur mainly in non-narrative contexts (without a temporal sequencing role). The
results show that in 23 % of the cases, the IMP has narrative usages boosted by a
subjective reading and accompanied by a temporal adverbial that increments the
reference point R. The consequence is the forward temporal interpretation, as in
example (22). Sentences in (23), (24) and (25) are the translation into FR, IT and
RO of (20) taken from the multilingual translation corpus described in Grisot and
(22) Alice was beginning to get very tired of sitting by her sister on the bank, and
of having nothing to do. Once or twice she had peeped into the book her sister
was reading, but it had no pictures or conversations in it, “and what is the use
of a book,”thought Alice, “without pictures or conversations?”(Literature
Corpus, Alice’s adventures in Wonderland, L. Carroll)
(23) Alice commençait à se sentir très lasse de rester assise à côté de sa sœur, sur le
talus, et de n’avoir rien à faire: une fois ou deux, elle avait jeté un coup d’œil
sur le livre que lisait sa sœur; mais il ne contenait ni images ni dialogues: Et,
pensait Alice, “à quoi peut bien servir un livre où il n’y a ni images ni
(24) Alice cominciava a sentirsi mortalmente stanca di sedere sul poggio, accanto a
sua sorella, senza far nulla: una o due volte aveva gittato lo sguardo sul libro
che leggeva sua sorella, ma non c’erano imagini nè dialoghi, “e a che serve un
libro,”pensò Alice, “senza imagini e dialoghi?”
(25) Alice începuse săse simtăfoarte obosităstând pe bancălângăsora ei şi
neavând nimic de făcut: o datăsau de douăori trase cu ochiul la cartea pe care
sora ei o citea dar nu avea poze sau dialoguri „Şi care e rostul unei cărţi”se
gândi Alice, “fărăpoze sau dialoguri?”.
These examples show that multilingual translation corpora reveal information
that is hidden in a monolingual or even bilingual analysis. The EN SP from (20) has
a narrative usage (forward temporal inference) and is translated into IT and RO
through tenses with a preterit form: passato remoto and perfectul simplu respec-
tively. The situation is different for FR, where the SP is translated through an
IMP. A close look at the FR sentence allows us to identify that Alice’s subjective
368 C. Grisot
perspective on the event of thinking and the discourse connective et (“and”)
accompany the IMP. In this sentence we are dealing with a narrative IMP (“im-
parfait de rupture”) that becomes uncovered in a multilingual contrastive analysis.
Experiment 2 was designed to test the narrativity feature for the SP based on
cross-linguistic transfer of semantic and pragmatic features. Two English native
speakers participated in this experiment and were instructed to judge 458 sentences
randomly chosen from the EN part of the corpus regarding the [±narrative] feature.
The two possible values of feature were explained and illustrated with several
examples. The results showed that annotators agreed on the label for 325 sentences
(71 %) and disagreed for 133 sentences (29 %). Disagreements were resolved in a
second round of the annotation experiment, where annotators were asked to insert
the connectives corresponding to the temporal and/or causal relation existent
between the two eventualities considered in order to test if the narrative or
non-narrative label was appropriate. Annotators disagreed only on 4 sentences.
These two experiments show that the [±narrative] feature, consisting of the
possible temporal and causal relations described in Fig. 1(Moeschler 2000) is cross
linguistically valid (i.e., it has been validated for FR, EN, IT and RO as discussed in
Grisot 2015). In order to be use this feature for MT, a much larger amount of
annotated data is needed. Because human annotation is costly, the labels should be
given automatically. This is the task of an automatic classiﬁer (described in Grisot
and Meyer 2014). The automatic classiﬁer was trained on the human annotated
corpus. When labelling an unknown corpus, the classiﬁer had similar results to
Aﬁrst run of an SMT system, which uses the classiﬁer trained on the annotated
data with the [±narrativity] feature, had slightly better results than without this
pragmatic feature. When trained and tested on automatically annotated data, the
[±narrativity] feature improves translation by about 0.2 BLEU points.
evaluation shows that verb tense translation and verb choice are improved by
respectively 9.7 % and 3.4 % (absolute), leading to an overall improvement of verb
translation of 17 % (relative) (for more detailed results see Meyer et al. 2013).
5 Discussion and Conclusion
Jacques Moeschler’s predictive model for temporal coherence in discourse is a
complex model consisting of several factors: lexical information (e.g. push–fall),
procedural contents of verb tenses and of connectives and contextual assumptions.
He suggests that these factors do not have the same status, but they are in a
hierarchy. Speciﬁcally, he proposed the following hierarchy:
BLEU (Bilingual Evaluation Understudy) is an evaluation measure for machine-translated texts.
It calculates the degree of resemblance to a human-translated text and it is a number between 0 and
1, where values closer to 1 represent more similar texts.
Temporal Coherence in Discourse: Theory …369
(27) Contextual assumptions >> procedural information of connectives >> pro-
cedural information of tenses >> conceptual information of verbs
The Model of Directional Inferences (MDI) proposed in Moeschler (2000,2002)
uses this hierarchy. He suggested that verbs (lexical information) and tenses convey
weak directional features while discourse connectives and contextual assumptions
convey strong directional features. This hierarchy is useful in case of conﬂicting
contextual and linguistic information.
I have shown in this paper the role played by one type of procedural information
encoded by verb tenses, the [±narrativity] feature, for temporal interpretation of a
discourse. This feature was empirically validated by human and automatic systems
in a multilingual translation corpus. Moeschler’s MDI model predicts that temporal
coherence in discourse depends not only on temporality (discourse temporal rela-
tions) but also on causality (discourse causal relations). I have shown that this
information is useful for improving the results of SMT systems in terms of the
choice of the verb and of the tense.
Finally, Grisot (2015) provides a more complex model for temporal coherence in
discourse that includes grammatical and lexical aspect, tense, temporal adverbials
and temporal connectives, as well as other linguistic and non-linguistic cues.
is needed to test each factor empirically, as well as their inter-
action with the help of complex statistical models.
Asher N (1993) Reference to abstract objects in discourse. Kluwer Academic Publishers,
Asher N, Lascarides A (2003) Logics of conversation. Cambridge University Press, Cambridge
Ahern A, Leonetti M (2004) The Spanish subjunctive: procedural semantics and pragmatic
inference. In: Pragmatics and beyond, vol 123, pp 35–56
Amenós-Pons J (2010) Telling a story in French and in Spanish: past tenses and temporal relations
revisited. Pragmatic perspectives on language and linguistics I: speech actions in theory and
applied studies, pp 395–424
Amenos-Pons J (2011) Cross-linguistic variation in procedural expressions: semantics and
pragmatics. In: Escandell-Vidal V, Leonetti M, Ahern A (eds) Procedural meaning: problems
and perspectives. Emerald Group Publishing Limited, Bingley, pp 235–266
Blakemore D (1987) Semantic constraints on relevance. Blackwell, Oxford
Blakemore D (2002) Relevance and linguistic meaning: the semantics and pragmatics of discourse
markers. Cambridge University Press, Cambridge
Carston R (1988) Implicature, explicature, and truth-theoretic semantics. In: Kempson R
(ed) Mental representations: the interface between language and reality. Cambridge University
Press, Cambridge, pp 155–181
Carston R (1993) Conjunction, explanation and relevance. Lingua 90(1):27–48
Carston R (2002) Thoughts and utterances: the pragmatics of explicit communication. Blackwell,
Some of these suggestions are investigated in (Grisot 2015).
370 C. Grisot
Cohen LJ (1971) The logical particles of natural languages. In: Bar-Hillel Y (ed) Pragmatics of
natural language. Reidel, Dordrecht, pp 50–68
de Saussure L (1997) Passé simple et encapsulation d’événements. Cahiers de linguistique
de Saussure L (2000) Quand le temps ne progresse pas avec le passé simple. Cahiers Chronos
de Saussure L (2003) Temps et pertinence: éléments de pragmatique cognitive du temps. De
Gazdar G (1979) Pragmatics: Implicature, presupposition, and logical form. Academic Press, New
Grice HP (1975) Logic and conversation. In: Cole P, Morgan J (eds) Speech acts, pp 41–58
Grice HP (1967) Presupposition and conversational implicature. In: Cole P (ed) Radical
pragmatics. Academic Press, New York, pp 183–198
Grice HP (1967) Logic and conversation: The william james lectures. Harvard University, MS
Grice HP (1989) Study in the way of words. Cambridge University Press, Cambridge
Grimes JE (1975) The thread of discourse. Walter de Gruyter, Berlin
Grisot C (2015) Temporal reference: empirical and theoretical perspectives. Dissertation,
University of Geneva
Grisot C, Cartoni B (2012) Une description bilingue des temps verbaux: étude contrastive en
corpus. Cahiers de linguistique française 30:101–117
Grisot C, Costagliola M (2014) What do translation corpora tell us about the semantics and
pragmatics of tense? The case of the English preterit. Chronos 11, 16–18 Jun 2014. University
of Pise, Italy
Grisot C, Moeschler J (2014) How do empirical methods interact with theoretical pragmatics? The
conceptual and procedural contents of the English Simple Past and its translation into French.
In: Romero-Trillo J (ed) Yearbook of corpus linguistics and pragmatics 2014: new empirical
and theoretical paradigms. Springer International Publishing, Switzerland, pp 7–33
Grisot C, Meyer T (2014) Cross-linguistic manual and automatic annotation for a pragmatic
feature of verb tense. In: Proceedings of LREC, 28–30 May 2014, Reykjavik, Island
Grosz BJ, Sidner CL (1986) Attention, intentions, and the structure of discourse. Comput Linguist
Halliday M, Hasan R (1976) Cohesion in English. Longman, London
Hobbs J (1979) Coherence and coreference. Cogn Sci 3:67–90
Hobbs JR (1985) On the coherence and structure of discourse. Center for the study of language and
information. Stanford University, pp 85–37
Horn L (1984) Toward a new taxonomy for pragmatic inference: Q-based and R-based
implicature. In: Schiffrin D (ed) Meaning, form, and use in context. Georgetown University
Press, Washington DC, pp 11–42
Horn L (1992) The said and the unsaid. Ohio State University Working Papers in Linguistics.
Horn L (2004) Implicature. In: Horn L, Ward G (eds) The handbook of pragmatics. Blackwell,
Oxford, pp 3–28
Horn LR (2006) The border wars: a neo-gricean perspective. In: von Heusinger K, Turner K
(eds) Where semantics meets pragmatics: the Michigan papers. Elsevier, Amsterdam, pp 21–48
Horn LR (2007) Toward a fregean pragmatics: Voraussetzung, nebengedanke, andeutung. In:
Kecskes I, Horn L (eds) Explorations in pragmatics: linguistic, cognitive and intercultural
aspects. Mouton de Gruyter, Berlin, pp 39–69
Kamp H, Reyle U (1993) From discourse to logic: introduction to modeltheoretic semantics of
natural language, formal logic and discourse representation. Kluwer Academic Publishers,
Kamp H, Rohrer C (1983) Tense in texts. In: Bauerle R, Schwarze C, von Stechow A
(eds) Meaning, use and interpretation of language. Walter de Gruyter, Berlin, pp 250–269
Kehler A (2004) Discourse coherence. In: Horn L, Ward G (eds) The handbook of pragmatics.
Blackwell, Oxford, pp 241–265
Temporal Coherence in Discourse: Theory …371
Labov W, Waletzky J (1967) Narrative analysis: oral versions of personal experience. In: Helm J
(ed) Essays on the verbal and visual arts: proceedings of the 1966 annual spring meeting of the
American Ethnological Society. University of Washington Press, Seattle, pp 12–44
Lascarides A, Asher N (1993) Temporal interpretation, discourse relations and common-sense
entailment. Linguist Philos 16:437–493
Leonetti M, Escandell-Vidal V (2003) On the quotative readings of Spanish imperfect. Cuaderons
Levinson SC (1983) Pragmatics. Cambridge textbooks in linguistics, Cambridge
Levinson SC (2000) Presumptive meanings: the theory of generalized conversational implicature.
MIT Press, Cambridge, MA
Longacre R (1983) Vertical threads of cohesion in discourse. In: Neubauer F (ed) Coherence in
natural-language texts. Helmut Buske Verlag, Hambourg, pp 99–113
Mann WC, Thompson SA (1986) Relational propositions in discourse. Discourse Process
Mann WC, Thompson SA (1988) Rhetorical structure theory: toward a functional theory of text
organization. Text 8(3):243–281
Meyer T, Grisot C, Popescu-Belis A (2013) Detecting narrativity to improve English/French
translation of simple past verbs. In: Proceedings of DiscoMT 2013 (ACL workshop on
discourse in machine translation), Soﬁa, Bulgaria, pp 33–42
Moeschler J (1985) Argumentation et conversation: éléments pour une analyse pragmatique du
discours. Hatier, Paris
Moeschler J (1993) Aspects pragmatiques de la référence temporelle: indétermination, ordre
temporel et inference. Langages 112:39–54
Moeschler J (1994) Connaissances d’arrière-plan et accessibilité contextuelle dans l’interaction
verbale. In: Trognon A et al (eds) La construction interactive du quotidien. Presses
Universitaires de Nancy, Nancy, pp 69–73
Moeschler J (1996) Théorie pragmatique et pragmatique conversationnelle. Armand Colin, Paris
Moeschler J (2000) L’ordre temporel est-il naturel? Narration, causalité et temps verbaux. In:
Moeschler J, Béguelin M-J (eds) Référence temporelle et nominale. Peter Lang, Berne,
Moeschler J (2001) The Geneva school. In: Brinker K, Antos G, Heinnemann W, Sager SF
(eds) Linguistics of text and conversation. An international handbook of contemporary
research. Walter de Gruyter, Berlin, pp 952–957
Moeschler J (2002) Economy and pragmatic optimality: the case of directional inferences. Gener
Grammar Geneva 3:1–20
Moeschler J (2005) Connecteurs pragmatiques, inférences directionnelles et représentations
mentales. In: Molendijk A, Vet C (eds) Temporalité et attitude: structuration du discours et
expression de la modalité. Cahiers Chronos 12, pp 35–50
Moeschler J, Grisot C, Cartoni B (2012) Jusqu’où les temps verbaux sont-ils procéduraux?
Nouveaux Cahiers de linguistique française 30:119–139
Moeschler JH, Jayez, J-M, Luscher, de Saussure LL, Sthioul B, Kozłowska M (1998) Le Temps
des événements: Pragmatique de la référence temporelle. Kimé, Paris
Nicolle S (1997) Conceptual and procedural encoding: criteria for the identiﬁcation of
linguistically encoded procedural information. In: Groefsema M (ed) Proceedings of University
of Hertfordshire relevance theory workshop. Peter Thomas and Associates, Chelmsford,
Nicolle S (1998) A relevance theory perspective on grammaticalization. Cogn Linguist 9:1–36
Reboul A, Moeschler J (1998) Pragmatique du discours. De l’interprétation de l’énoncé à
l’interprétation du discours. Armand Colin, Paris
Reichenbach H (1947) Symbolic logic. University of California, Berkeley
Roulet E (1997) A modular approach to discourse structures. Pragmatics 7:125–146
Roulet E, Auchlin A, Moeschler J, Rubattel C, Schelling M (1985) L’articulation du discours en
français contemporain. Peter Lang, Berne
372 C. Grisot
Samaradzic T (2013) Dynamics, causation, duration in the predicate-argument structure of verbs: a
computational approach based on parallel corpora. Dissertation, University of Geneva
Sanders T (1997) Semantic and pragmatic sources of coherence: on the categorization of
coherence relations in context. Discourse Process 24:119–147
Sanders T (2005) Coherence, causality and cognitive complexity in discourse. In:
Proceedings/Actes SEM-05, ﬁrst international symposium on the exploration and modelling
of meaning, pp 105–114
Sanders T, Spooren W, Noordman L (1992) Toward a taxonomy of coherence relations. Discourse
Sanders T, Noordman L (2000) The role of coherence relations and their linguistic markers in text
processing. Discourse Process 29:37–60
de Saussure L (2011) On some methodological issues in the conceptual/procedural distinction. In:
Escandell-Vidal V, Leonetti M, Ahern A (eds) Procedural meaning: problems and perspectives.
Emerald Group Publishing Limited, Bingley, pp 55–79
de Saussure L, Sthioul B (1999) L’imparfait narratif: point de vue (et images du monde). Cahiers
de praxématique 32:167–188
de Saussure L, Sthioul B (2005) Imparfait et enrichissement pragmatique. Cahiers Chronos
de Saussure L, Morency P (2012) A cognitive-pragmatic view of the French epistemic future.
J French Lang Stud 22:207–223
Smith N (1990) Observations on the pragmatics of tense. UCL working papers in linguistics,
Smith CS (2003) Modes of discourse: the local structure of texts. Cambridge University Press,
Sperber D, Wilson D (1986/1995) Relevance: communication and cognition. Blackwell, Oxford
Tahara I (2000) Le passé simple et la subjectivité. Cahiers de linguistique française 22:189–218
Vendler Z (1957) Linguistics in philosophy. Cornell University Press, Ithaca
Wilson D, Sperber D (1993) Linguistic form and relevance. Lingua 90:1–25
Wilson D, Sperber D (1998) Pragmatics and time. In: Carston R, Uchida S (eds) Relevance theory:
applications and implications. John Benjamins, Amsterdam, pp 1–22
Temporal Coherence in Discourse: Theory …373