From Conceptual Time to Linguistic Time 
Michel Gagnon* 
Machina Sapiens 
Guy Lapalme t 
Universit4 de Montr4al 
In this paper, we present a method for generating French texts conveying temporal information 
that integrates Discourse Representation Theory (DRT) and Systemic Grammar Theory. DRT 
is used to represent temporal information and an intermediate semantic level for the temporal 
localization expressed by temporal adverbial phrases and verb phrases. This representation is 
then translated into a syntactic form using Systemic Grammar Theory. We have implemented 
this method in a working prototype called Prdtexte. 
1. Introduction 
In speaker-generated texts, reference is made to facts taking place in time. To use the 
same kind of references in automatically generated text, the mechanisms that govern 
the expression of temporal concepts must be identified. There is no simple or direct 
mapping between conceptual time, as it is perceived in the real world, and linguistic 
time, which refers to the way time is formulated in language. There may be different 
ways to present the same temporal concept in a text, and a single linguistic marker 
can be used to convey different temporal meanings. 
For example, the discourse below (Discourse 1) is a text generated by Pr4texte, 
a system we developed for implementing the expression of temporal localization in 
French texts 1. It is a slightly modified version of an example used by Bras (1990) for the 
extraction of temporal information in text analysis. The sentences report occurrences 
that are facts taking place in time. We have inserted labels in parentheses to distinguish 
the twelve occurrences reported in the text. 
Hier l'avion a effectu6 un vol (ol). A 8h00 
il a quitt6 Paris (02). Quand il a survol6 
Barcelone (o3), le r6acteur fonctionnait (o4). 
~, 10h15, un voyant a clignot6 (o5). Aupar- 
avant, il s'6tait allure6 (o6). Puis il s'6tait 
6teint (o7). Pendant 35 minutes, l'avion a 
survol6 lamer (o8). Puis il a atteint la c6te 
(o9). Jusqu'a 10h50, il a survol6 l'Alg6rie 
(o10). A llh30 il 6tait sur la piste (o11). A 
ce moment-la le r6acteur a explos6 (o12). 
Yesterday the plane made a flight (ol). At 
8:00 A.M. it left Paris (o2). When it flew over 
Barcelona (o3), the engine was working (o4). At 
10:15, a warning light flashed (o5). Previously it 
had come on (o6). Then it had gone out (o7). For 
35 minutes the plane flew over the sea (o8). Then 
it reached the coast (o9). Until 10:50 it flew over 
Algeria (olo). At 11:30 it was on the landing run- 
way (o11). At this moment the engine exploded 
(o12). 
Discourse 1 
* 3535 Queen-Mary, Bur. 420, Montr6al (Quebec), Canada H3V 1H8, Tel: (514) 733-3959. E-maih 
gagnon@iro.umontreal.ca. This article was written while the author was at "Laboratoire Langue, 
Raisonnement et Calcul" of IRIT, Toulouse, France. J- D6partement cl'informatique et de recherche op6rationnelle, C. P. 6128, Succ. Centre-Ville, Montr6al 
(Qu6bec), Canada H3C 3J7. E-mail: lapalme@iro.umontreal.ca. 1 The French text shown on the left in Discourse 1 was generated by Pr6texte; we give an English 
translation on the right. 
@ 1996 Association for Computational Linguistics 
Computational Linguistics Volume 22, Number 1 
In Discourse 1, we find two types of temporal markers: verb tense and what we 
call adverbials of temporal location (ATL). An ATL is an adjunct, such as yesterday, 
until 10:50, or when it flew over Barcelona, that provides information about the temporal 
localization of an occurrence or its duration, or both at the same time. 
For verb tense, we distinguish different ways of indicating localization in the past. 
Three French verb tenses can be used: pass~ composG imparfait, and plus-que-parfait; their 
closest equivalents in English are the simple past, past progressive, and past perfect. 
The passf compos~ il a survold 'it flew over' presents the occurrence as an event and 
localizes it in relation to the time of speech. With the plus-que-parfait il s'dtait allum~ 'it 
had come on', the occurrence is also presented as an event, but localized in relation 
to a perspective point other than speech time. The imparfait le r~acteur fonctionnait 'the 
engine was working' presents the occurrence as being in progress. For present and 
future tenses there are fewer options than for past tense, but more than one form is 
available for theses tenses as well. 
For ATLs, temporal localization can be achieved in many ways; for example, in 
relation to the time of speech (hier 'yesterday'), by designating an absolute temporal 
location (~ 8hO0 'at 8:00 A.M.'), or in relation to another fact (puis 'then', ~ ce moment-ld 
'at this moment', quand il a survold Barcelone 'when it flew over Barcelona'). To this 
variety in the semantics of localization we must add the variety of syntactic forms. 
Localization can be expressed by an adverb (puis 'then'), a prepositional phrase (jusqu'd 
10h50 'until 10:50'), a nominal phrase (le lendemain 'the day after') or a subordinate 
clause (quand il a survold Barcelone 'when it flew over Barcelona'). 
No text generator has yet been developed to solve the problem of the expression 
of time. The ones that have tackled this problem have focused on the production of 
verb tenses, without solving the choice of temporal adverbs. The work presented in 
this paper addresses the problem of generating the elements that convey temporal 
localization in French, including both verbs and temporal adverbs. 
In a previous paper (Gagnon and Lapalme 1992), we proposed a method of inte- 
grating the expression of temporal concepts into the text-generation process. In par- 
ticular, we showed how to produce different types of text in French from a single 
representation of events. Unfortunately, the method governing the planning process 
was too determined by temporal concepts, so it was difficult to link this planning 
process with other frameworks, such as the schema proposed by McKeown (1985) or 
Rhetorical Structure Theory (RST) (Mann 1991; Hovy 1991). 
As we were not really successful in integrating the expression of time in French 
into the text-generation process, we decided to pursue our research with a different 
perspective. We designed a system covering many of the possibilities of expressing 
time in French, our hypothesis being that the achievement of this task would facilitate 
the design of a text-planning process. We believed it would be easier to organize the 
structure of the discourse with a better understanding of the way temporal information 
can be expressed by adverbs and verb tenses. We started from the work of Bras (1990), 
who proposes a method of extracting the temporal structure of a text, according to 
Discourse Representation Theory (Kamp 1981), that relies on an analysis of adverbials 
of temporal location made by Molin6s (1990). 
To implement the production of ATLs and verb tenses, we have chosen Systemic 
Grammar (Halliday 1985; Berry 1975, 1976), which formulates the syntactic structure 
of a sentence as the result of a sequence of semantic choices. We developed a grammar 
interpreter inspired by Nigel (Matthiessen and Bateman 1991), but departing from it 
in many respects; in particular, our representation of the production of verb tenses 
and adverbs is quite different. 
In this paper, we discuss the elements required to produce a text such as Dis- 
92 
Gagnon and Lapalme From Conceptual Time to Linguistic Time 
/ 
DEEP / 
GENERATIONN 
f 
SURFACE GENERATION", 
Conceptual 
representation 
Discourse 
representation 
t 
Semantic 
representation 
1 
Syntactic 
representation 
Occurrences as primitive concepts 
Temporal relations 
The tempolral Jocaligafion is represented as an overlapmg relation 
Sellmentafion of the conceptual ret~resentaaon 
Structured information 
Rhetorical relations 
Linearization 
Choice of aspect 
Identification of temporal markers 
Temporal adverbial Verb tense 
Figure 1 
The global process. 
course 1. The process starts from a conceptual representation that encodes the facts to 
be reported in the text, associated with their position in time. The information at this 
objective conceptual level must be translated into a semantic representation where the 
facts are presented according to a subjective perspective. The semantic representation 
is then used to produce the text. We have concentrated our attention on this last stage, 
but we cannot avoid the problem of determining how the representation used at this 
level is obtained from previous levels. In the following sections, we describe the two 
stages of the text-generation process. 
2. The Global Process 
It is generally accepted that the generation process requires at least two parts. The 
first part, deep generation, is a planning process in which the content and the overall 
structure of the text are established. In the second part, surface generation, the words 
and the syntactic structure of the text are chosen. 
Figure 1 summarizes our view of the global process, starting from a conceptual 
representation that contains occurrences and relations between them. The fact that 
an occurrence takes place at a certain time is expressed by an overlapping relation 
between this occurrence and the object representing this time. 
The deep generation process is decomposed into two steps. In the first step, the 
conceptual representation is segmented and structured to build a discourse represen- 
tation. In our discourse representation, which uses Segmented Discourse Representa- 
93 
Computational Linguistics Volume 22, Number 1 
tion Theory (SDRT) (Asher 1993), the information is cut into smaller segments each 
of which contains the information to be expressed by a single sentence. The structure 
linking these segments relies on a set of rhetorical relations. 
In the second step of the deep generation process, the discourse representation 
is traversed and, for each segment, rules are applied to identify the feature values 
needed to translate it into a sentence. We thus obtain a linear structure in which each 
element is a set of features that determine the syntactic form of the sentence. 
In the surface generation process, the information in the semantic representation 
is used to select the appropriate syntactic structure for the expression of time: an 
adverbial of temporal location (ATL) or a verb phrase (VP), or both. 
3. The Deep Generation Process 
Although our work focuses on surface generation, we cannot ignore the issue of deep 
generation, because the nature of the semantic representation is determined not only 
by the syntax of the language, but also by the temporal concepts available. Therefore, 
in this section, we first present the conceptual representation that induces the semantic 
representation used by our generator. We then explain the intermediate discourse level. 
We do not know yet in detail how to produce the semantic representation from the 
conceptual representation, but we do have an idea of what information each level 
of representation must contain and what choices must be made at each stage of the 
process. 
3.1 Conceptual Representation 
To represent temporal concepts in Pr6texte, we chose the principles of Discourse Rep- 
resentation Theory (DRT), which offers one of the most interesting explanations of 
how temporal notions are conveyed by a text. DRT was developed to deal with spe- 
cific problems of discourse understanding: in particular, problems with anaphora and 
the differences between some verb tenses, with respect to temporal localization. Our 
goal is not to show how this theory can be used for generation but rather to use its 
principles as a convenient formalism for the representation of time. 
In DRT, a text is associated with a Discourse Representation Structure (DRS) that 
is updated incrementally by the processing of each sentence. A DRS is a structure 
containing a set of entities and a set of conditions on these entities. There are different 
types of entities in DRT: 
a temporal fact can either be presented as an event (having a punctual 
aspect), or as a situation (having a certain extent in time, but considered 
from an internal perspective at a given moment in time); 
a temporal constant that designates a segment of the temporal axis; 
entities that participate in the events or situations. 
In Pr6texte, conceptual knowledge is represented as a DRS, which we adapted slightly 
for text generation. We do not distinguish between events and situations in the concep- 
tual representation, because we want this level to remain independent of the language. 
Furthermore, we feel that this distinction should not be encoded at the conceptual 
level, rather the generation system should choose among these possibilities. There- 
fore, at the conceptual level we use the concept of occurrence for either an event or a 
situation. 
94 
Gagnon and Lapalme From Conceptual Time to Linguistic Time 
n tl t2 t3 t4 01 02 03 04 a l p b r 
plane(a) 
flight(l) 
engine(r) 
city(p) 
Paris(p) 
city(b) o20 t2 
Barcelona(b) ol < n 
o1: make(a, 1) o3 < n 
o 2" leave(a,p) o~ C tl 
o3: flightover(a, b) o4 < n 
04: work(r) o3 C 04 
n c t3 03 C) t4 
tl = Sept. 10 1992 
t2 = Sept. 10 1992 at 8:10 am 
t3 = Sept. 11 1992 
t4 = Sept. 10 1992 at 9:00 am 
02 <n 
Figure 2 
Conceptual representation for the first three sentences of Discourse 1. 
There are essentially two ways of considering time or, to be more precise, the 
notion of temporal location: either temporal location is determined using a preexisting 
time scale, or it is deduced from the occurrence. Following Kamp (1981), we think that 
the second possibility, in which temporal location is a relative concept, is more suitable 
for natural text processing. Treating occurrences as entities, rather than making them 
subordinate to temporal intervals or points, has been proposed by Davidson (1967). 
An occurrence may be represented in relation to another temporal object, without 
any reference to its own location in time. This approach eases the representation of 
underspecified temporal localizers--an important point for our semantics. For further 
discussion of the advantages of taking occurrences as primitives, see (Bras 1990; Kamp 
1979, 1981). 
In the conceptual representation, we find four types of information: 
• the description of occurrences; 
• the description of participants in the occurrences; 
• the description of temporal localizers, which are called temporal 
constants (they usually refer to time periods of the calendar); 
• temporal relations between occurrences and temporal localizers: 
The relation < represents temporal precedence. 
The overlap relation C) indicates that two temporal objects are 
somehow simultaneous. Thus, in our representation, "Y happens 
at time X" is represented by "Y temporally overlaps X". 
The relation C expresses the fact that the temporal extent of a 
temporal object is a subset of the temporal extent of another 
object. 
Figure 2 shows the part of the DRS used to generate the first three sentences of 
Discourse 1. It contains five temporal constants: n, tl, t2, t3 and t4. It is not clear how 
95 
Computational Linguistics Volume 22, Number 1 
these temporal constants are to be described in DRT, so we have proposed elsewhere 
a formalization of the type of objects designated by these constants (Gagnon and Bras 
1994). In this article, we give only an English description of them: n represents speech 
time, which, in Figure 2, is included in the time represented by t3 (September 11 1992). 
Four occurrences are represented: ol, o2, o3, and o4, all of which take place before n. 
Not all temporal relations in the DRS need to be given as input because many 
relations can be inferred using three kinds of knowledge: 
• The representation of conventional time to identify a specific period in 
time; this representation relies on a structure of conventional time, 
together with reasoning mechanisms to deduce temporal relations (see 
Gagnon and Bras \[1994\] for an implementation of such a structure). For 
example, from this knowledge we can deduce that September 10 must be 
before September 11, which would be represented as tl < t3. Similarly, we 
can deduce t2 C tl, t4 C tl, and t2 < t4. 
• World knowledge about the occurrences: knowing that o2, 03, and o4 are 
part of Ol implies that they are all temporally included in it. 
• A reasoning mechanism on the temporal relations, using a set of axioms, 
such as: 
Vx, y(xOyvx < yvy < x) 
Vu, v,x,y(uOxAvOyAx < y~ uOvVu < v) 
The first axiom states that for any two times, either they overlap or one 
precedes the other. The second axiom states that if two other times u, v 
overlap two times x, y that are in a precedence relation, either u overlaps 
v, or it precedes it. 
So in Figure 2, from the relations 02 O t2, 03 O t4, and t2 < t4, we can infer 
o2 < 03 V o20 o3. From world knowledge, we can infer that o2 and 03 cannot overlap 
(leaving of Paris cannot overlap flying over Barcelona); Therefore, we conclude that 
02 ~ 03. 
3.2 The Discourse Representation 
To generate a text from an input such as Figure 2, we must choose a discourse structure 
that segments the message into sentences. 
Figure 3 illustrates one discourse representation, inspired by the Segmented Dis- 
course Representation Theory (SDRT) proposed by Asher (1993), which extends DRT 
by adding rhetorical relations such as those found in RST (Mann and Thompson 1987). 
A discourse structure that contains the same information as in Figure 2, except that 
it has been segmented, we call a Segmented Discourse Representation Structure (or 
SDRS). The top-level DRS contains three small DRSs that are linked by rhetorical rela- 
tions: each DRS corresponds to a sentence. In addition to these three small DRSs, the 
top-level DRS contains the global text information: the description of participants and 
the description of speech time. We do not yet produce this discourse structure, but 
we are working on this problem, using the results of researchers who have applied 
SDRT to the analysis process (Lascarides and Asher 1993; Bras and Asher 1994). In 
the discourse structure of Figure 3, one sentence is elaborated by two other sentences 
that constitute a narration. 
96 
Gagnon and Lapalme From Conceptual Time to Linguistic Time 
plane(a) 
fight(l) 
city(p) 
Paris(p) 
n t3 a l p b r 
Ol tl 
o1: make(a, 1) 
tl = Sept. 10 1992 
Ol Kn 
Ol C tl 
Elaboration 
Barcelona(b) 
engine(r) 
t3 = Sept. 11 1992 
n C t3 city(b) 
02 t2 
o2: leave(a, p) 
t2 = Sept. 10 1992 at 8:10 am 
02 Kn 
02 0 t2 
Narration 
03 04 t4 
03: flightover(a,b) 
04: work(r) 
O3 (n 
04 <n 
t4 = Sept. 10 1992 at 9:00 am 
03 C 04 
03 0 t4 
Figure 3 
Discourse representation for the first three sentences of Discourse 1. 
3.3 Semantic Representation 
The discourse structure is then translated into a semantic representation of the form 
$1, $2,..., Sn where Si designates the i th element of a semantic representation S. Transla- 
tion of the SDRS is obtained by a depth-first traversal of the DRSs it contains. For each 
DRS, we establish its corresponding feature structure in the semantic representation. 
Figure 4 is a semantic representation produced from the SDRS of Figure 3. Each 
structure contains five features. The feature message refers to the occurrence that must 
be reported by the sentence, and specifies its aspect. We distinguish, as Kamp does, 
two aspects that can be used to present an occurrence: event or situation. Situations 
can be open or resulting. A situation is open when the speaker/hearer is located at a 
time within an occurrence. A resulting situation is the state following the termination 
of an occurrence. In French, the event aspect for a past occurrence results in the use 
of the pass~ composd (simple past in English). The imparfait and the plus-que-parfait 
correspond to open and resulting situations (the closest tenses in English are the past 
progressive and the past perfect). In the first two elements of Figure 4, the occurrence 
is presented as an event, whereas in the last one it is presented as an open situation. 
Among the four occurrences contained in the DRS, only three of them constitute the 
main "message" of the text: ol, o2, and o4. 
The four other features in a structure Si give the value of four temporal markers 
that express the localization of the occurrence. These markers correspond to the four 
markers proposed by Kamp and Rohrer (Bras 1990) for the analysis of texts, which we 
have adapted for text generation. They can be seen as an extension of Reichenbach's 
markers (1947). Essentially, the values of these four features depend on two data: 
• the DRSs to which the visited DRS is attached in the discourse structure, 
and 
• the rhetorical relations. 
97 
Computational Linguistics Volume 22, Number 1 
Figure 4 
message = 
event(ol) 
N=n 
REF = nil 
PERSP = n 
LOC = tl 
S1 
message = 
event(o2) 
N=n 
R =Ol 
PERSP = n 
LOG = t2 
$2 
message : 
open_situation ( o 4 ) 
N=n 
R=o2 
PERSP = n 
LOC = o3 
S3 
Semantic representation for the first three sentences of Discourse 1. 
O1 : .... 
~. Elaboration 
102 .... LNa a onL04 .... INaati°nIO5 .... I a ati°n 
~. Flashback 
h 08: .... I 
06: .... \[ Narration ~, 07: .... \] 
Figure 5 
Discourse representation for the first seven sentences of Discourse 1. 
The marker N represents the time of speech. This marker is constant in our exam- 
ple, but it could be locally altered in the discourse, in indirect speech for example. We 
did not study such cases, but we think that the marker N would be required to deal 
with them. 
Perspective point PERSP refers to an instant from which the occurrence must be 
considered. Usually it is the same as the time of speech, but in some cases, such as a 
flashback, it has a different value. In Discourse 1, there is one such case. The fifth and 
sixth sentences (where occurrences o6 and o7 are reported) constitute a flashback: the 
perspective point is the occurrence of the fourth sentence (os). In discourse structure, 
the flashback is represented with a rhetorical structure. Consider for example the 
discourse structure for the first seven sentences of Discourse 1 as sketched in Figure 5. 
For the translation of the two DRSs containing 06 and o7, the value of PERSP will be 
the occurrence o5, since the DRS containing this occurrence dominates the two others 
with the relation flashback. For the next DRS, the one containing os, the perspective 
point will be reset to the value it had before entering the flashback, that is, the value 
when the DRS of o6 was considered. 
The value of PERSP is used for the choice of verb tense. In Discourse 1, the 
flashback results in the choice of the plus-que-parfait. 
LOC represents the temporal location of the occurrence reported. If this occurrence 
overlaps another temporal object, this object may be used as a value for LOC. In 
98 
Gagnon and Lapalme From Conceptual Time to Linguistic Time 
Figure 4, the values of LOC show that tl and t2 are used to localize the first two 
occurrences. In the third sentence, the situation corresponding to o4 is presented at 
the instant where o3 takes place. If no other temporal object overlaps the one that 
constitutes the message, the temporal location represented by LOC can be defined in 
relation to another temporal object. We will see examples of this in the next section. 
LOC represents the information to be expressed by an ATL in the sentence and does 
not necessarily have a value, because a sentence may not contain an ATL. 
In a text, when we want to express a succession of occurrences, we need a way 
to memorize the occurrence that is used as a reference for the localization of the 
next one. This is exactly the role of the marker REF. The values of REF are used to 
represent the progression of time in the discourse. Each time a sentence expresses a 
new temporal location (which can be an occurrence or a temporal constant), the value 
of REF is updated to this new value, and the temporal localization in the following 
sentence may be achieved in relation to this reference. The following rules are used 
for identifying the value of REF: 
1. Identify the S-antecedent, the DRS to which the current DRS is attached 
in the discourse structure, and Sa, the feature structure associated with 
this DRS in the semantic representation. 
2. If the occurrence reported as message in Sa is presented as a situation, it 
cannot be used as a reference point, since a situation cannot state a 
progression in time. 
3. If LOC has a value that is temporally more precise than the occurrence 
in the message, REF will take on its value, otherwise REF is bound to the 
occurrence in the message, if this occurrence is not presented as a 
situation. 
4. If LOC has no value and the occurrence in the message is a situation, the 
antecedent sentence does not state a progress in time. Therefore, REF 
takes the same value as in Sa. 
In Figure 4, the context for the first sentence is empty, so no value is given to 
REE For the second and third sentences, the value of REF is the event presented in 
the previous sentence. The occurrence in the third sentence is expressed as a situation, 
so it cannot be the reference for the fourth sentence (not shown in the figure). Con- 
sequently, REF for the fourth sentence takes the value of LOC in the structure of the 
third sentence: t4. We will see in Section 5 how the value of REF is used to produce 
the temporal adverb. 
The choice of aspect in building the semantic structure is achieved by taking into 
account pragmatic information and the interaction with other choices, such as the 
type of temporal localizer. Currently, we first identify the localizer that constrains the 
selection of aspect, but more study is needed to clarify their interaction. 
If an occurrence is presented as a situation, the temporal localizer must be a time 
included in it; but an event aspect cannot be combined with a localizer. In Figure 4, 
the occurrence of $1 must be presented as an event, since the localizer tl includes 
the occurrence. In $2, the occurrence is also an event, even if the localizer overlaps 
the occurrence: the overlapping relation does not prevent the existence of an inclusion 
relation. If an inclusion relation between t2 and o2 could be deduced, then the situation 
aspect could be chosen. In $3, the localizer is included in the occurrence of the message, 
so the situation aspect is selected. 
99 
Computational Linguistics Volume 22, Number 1 
message = 
event (oi) 
N=n 
R = nil 
P=n 
L = ch 
message = 
event(o2) 
N=n 
' R =01 
P=n 
L = ct2 
Figure 6 
An alternative semantic representation. 
message = 
event(03) 
N=n 
' R =02 
P=n 
L = ct4 
message = 
open_situation ( o 4 ) 
N=n 
' R =03 
P=n 
L = ct4 
) 
The semantic representation given in Figure 4 is not unique. Figure 6 shows an- 
other semantic representation built from the DRS of Figure 2. It contains a fourth 
sentence. The main difference from the previous representation is that 03, instead of 
acting as a localizer for o4, is the message of a sentence; t3, referring to a moment 
located two hours after t2, localizes o3. Therefore, instead of the third sentence of 
Discourse 1, we would obtain these two sentences: 
Deux heures plus tard, il a survol6 Two hours later, it flew over Barcelona (o'3). At 
Barcelone (o~). Ace moment-la, le r6acteur this moment, the engine was working (o'4). 
foncfionnait (o~). 
Once the semantic representation is produced, the adverbial or temporal location (ATL) 
and the verb phrase (VP) can be generated independently. The syntactic form of the 
verb phrase is determined by the combination of the following information: 
• temporal relation between localizer LOC and speech time N; 
• temporal relation between localizer LOC and perspective point PERSP; 
• aspect of the occurrence. 
The choice of the syntactic structure of the ATL depends on the value of LOC, 
which may refer to N or REF. The interaction of temporal information conveyed by 
verb tense and adverbs is taken into account in the process of translation from the con- 
ceptual representation to the semantic level, where the choices of aspect, perspective 
point PERSP and localizer LOC are made. 
We still have not entirely solved the problem of choosing among all semantic 
representations that can be built from a DRS. In the current implementation of Pr6texte, 
we have identified a set of rules to produce the semantic representation. In particular, 
these rules insure that the values of the four temporal markers are coherent with the 
aspect chosen to present the occurrences. What remains to be done, essentially, is to 
identify the knowledge that governs these rules causing them to select a particular 
semantic representation. 
3.4 Representation of Localization 
We have argued in the previous section that four temporal markers are needed to 
express the temporal location of an occurrence. In this section, we discuss the marker 
LOC, the localizer that provides information about the location in time of the occur- 
rence using another entity. The localizer is usually a temporal constant, but it can also 
be another occurrence, whose approximate location in time is already known. 
An ATL can convey localizers of two types: in the first type, a localizer directly 
identities the temporal zone of an occurrence using another temporal object that over- 
laps it; in the second type the temporal zone of an occurrence is conveyed in relation 
100 
Gagnon and Lapalme From Conceptual Time to Linguistic Time 
to another localizer. In Figure 4, all occurrences are localized directly. Occurrence 04 is 
localized directly by another occurrence, whereas 01 and o2 are localized directly by 
a temporal constant. For $1 and $2, the values for LOC are simply constants tl and 
t2. The same temporal localization can usually be expressed in many ways, and we 
must also specify how these constants are worded. For example, in Discourse 1, T1 
has been translated into hier 'yesterday' but it could also have been translated into le 
10 septembre 1992 'September 10th 1992', or mercredi dernier 'last Wednesday'. 
Thus, the value of LOC in the semantic representation determines the expression 
of the localization, giving rise to three main problems: 
• how to represent the temporal constants in the conceptual representation; 
• how to determine the link between these conceptual temporal constants 
and their semantic representation, which specifies how they are to be 
expressed in the text; and 
• how to implement the selection mechanism, which relies on pragmatic 
and stylistic information to choose between the many different ways of 
expressing the same temporal localization. 
In Gagnon and Bras (1994), we gave a solution to the first two problems but the 
last one still remains to be solved. Here, we discuss only the second problem: the 
semantic expression of localization. 
First, a few words about temporal context: usually, temporal localizers may be 
understood only in reference to some time in the context. For example, in on April 
15th, it is assumed that it is clear which year this time is part of. Thus, we take for 
granted that all expressions of temporal localization are made in relation to such a 
contextual time. 
Let ti be a temporal constant, taken from the conceptual representation, which is to 
be used as a localizer in the semantic representation, and tcont the contextual time. The 
expression in the semantic representation is based on a term of the following form: 
\[ti, Type, Naming\] 
The first argument is the identifier of the constant in the conceptual representation. 
The second argument is the type of the temporal localizer (day, month, year, etc.). Thus, 
tcont may be decomposed into times of type Type, of which t i is one. The last argument 
names the time referred to by the localizer ti. There is exactly one time in the "real 
world" that corresponds to the temporal constant ti. We call it objective time. We use 
the notation t* to represent the objective time that corresponds to a localizer ti. 
For example, the expression for the temporal localizer en avril 'in April', would be 
something like this (here t67 is the corresponding temporal constant in the conceptual 
representation): 
\[t67, mort th, april\] 
The naming april is not the syntactic word "April" but an internal keyword that 
helps distinguish between the time referred to and the other months of the contextual 
year. 
An important distinction is made between a temporal constant ti and a objective 
time t*. The constant ti pertains to the way a temporal location is expressed in the 
discourse, whereas t~ can be considered the corresponding portion of time in the real 
world. More than one temporal constant may correspond to a single objective time. 
101 
Computational Linguistics Volume 22, Number 1 
Table 1 
Relative localizers. 
Localizer Description 
inclin(\[ti, Ti, Ni\],\[tj, Tj, Nj\]) 
incl(\[ti, Ti, Ni\],\[tj, Tj, Nj\]) 
begin(\[ti, Ti, Ni\],\[tj, Tj, Nj\]) 
end(\[ti, Ti, Ni\],\[tj, Tj, Nj\]) 
after(\[ti, Ti, Ni\],\[tj, Tj, Nj\],D ) 
before(\[ti, T,, Ni\],\[tj, Tj, Nj\],D) 
relpos(X,\[ti, T, Ni\],\[tj, T, Nj\]) 
extent(\[ti, Ti, Ni\],\[G Tj, Nj\],\[tk, Tk, Nk\]) 
t~ is a time included in the time t 7 
t* is a time that includes the time tj* 
t~ is a time whose beginning overlaps the 
time t~ 
t~ is a time whose ending overlaps the time t; 
t~ is a time after the time t~ with a temporal 
distance D (expressed as a duration) 
t~ is a time before time t 7 with a temporal 
distance D (expressed as a duration) 
t~ is the X th (-X th, if X<0) time of type T 
after (before, if X<0) time tj* 
t~ is a time period starting at time t7 and 
ending at time t; 
Suppose, for example, that the discourse contains two temporal constants, correspond- 
ing to yesterday and two days after Robert's departure. We can imagine a situation where 
both designate exactly the same day. But it is not possible for a single temporal con- 
stant to correspond to more than one objective time. If it were, it would mean that 
an ATL could be ambiguous. In our computational perspective, we accept some un- 
derspecified ATLs (not precisely specifying the temporal location), but not ambiguous 
ones. 
A triplet is the simplest expression of a temporal localization. Usually, expressing a 
temporal localization is more complex, because the temporal constant used as localizer 
cannot be rendered directly in relation to the contextual time. This is the case, for 
example, if the localizer is a day, and the contextual time a year, because there is 
no natural way of decomposing a year into days 2. In these cases, we must express 
relations with some intermediate localizers, until we reach one that can be related to 
the contextual time. 
Let \[ti, Ti, Nj\] be the localizer and \[tj, Tj, Nj\], \[tk, Tk, Nk\] be intermediates localizers. 
Many kinds of relations can be distinguished. They are listed in Table 1. Note that 
these relations can be combined recursively. This means that the triplets used as argu- 
ments may also be represented using a relation. We will show examples of this in the 
following discussion. 
Among the arguments of these relations, one pertains to the temporal localizer, 
and one (two, in the case of the relation extent) pertains to an intermediate localizer 
to which the temporal localizer is related. We call this last argument an anchor, since 
it represents a time to which the relation must be "anchored" in order to deduce the 
time of localization. 
We will now give a short discussion of the relations of Table 1. In the follow- 
ing examples, hoe designates the temporal constant corresponding to the localizer of 
the occurrence, and n and trey designate the time of speech and the reference time, 
respectively. 
2 It is possible to name the day using the religious calendar, using something like the day of St-Andrew, 
but it is far from usual to do so, except maybe for holidays such as Christmas, Easter, or Thanksgiving. 
102 
Gagnon and Lapalme From Conceptual Time to Linguistic Time 
The first relation is the most frequent for expressing temporal localization. It is 
used to express localizations like le 3 avril 'on April 3rd'. In this case, the localizer 
could be formulated as: 
inclin ( \[ tto¢, day, 3\], \[t~, month, april\]) 
As expected, the intermediate localizer tl is to be interpreted in relation to the 
contextual year. The semantics of a more complex localizer like le matin du 3 avril 1994 
'on the morning of April 3rd 1994' is an example of using recursivity for the expression: 
inclin ( \[hoc, momen t-of-day, morning\], inclin ( \[ h, day, 3\], 
inclin (\[t2, month, april\], It3, year, 1994\] ))) 
The second relation, incl, is required to deal with adverbials such as le jour oft Paul 
est parti 'the day when Paul left', aujourd'hui 'today' and ce jour-l~ 'that day'. All of 
these refer to a day, but this day is not expressed in relation to a time that includes it. 
On the contrary, the other time is included in it: the time when Paul left in the first 
example, the time of speech in the second example, and the referent time associated 
with REF in the third one. Suppose that in the conceptual representation, 023 is the 
object representing the departure of Paul. These three examples could be represented, 
respectively, as: 
incl(\[hoc, day, _\], \[O23 , -, -\]) 
incl(\[ttoc, day, _\], \[n, _, _\]) 
incl ( \[ tloc, day, _\], \[ tref , -, -\] ) 
where "_" is used for arguments whose values are not relevant or unknown. 
The relations begin and end represent the case where only one boundary of the 
localizer is known. This results in an ATL such as depuis le 3 avril 'since April 3rd' or 
jusqu'au 3 avril 'until April 3rd' where meaning is ambiguous. What do we mean ex- 
actly, when we write that ti begins at time tj? That the initial boundary of ti is included 
in tj or that the ending of tj "meets" the beginning of ti? If the answer is that tj includes 
the initial boundary of ti, what is the constraint on the duration of ti? It is clear that it 
must be shorter than tj. We do not have any answers to these crucial questions, and 
other similar ones. These are problems that pertain to the deep generation process, 
which is not the focus of this paper. We think that at the level of the semantic repre- 
sentation, this relation need not be further clarified, since it corresponds to the way 
time is expressed in the language. In fact, all our relations have this underspecified 
nature. 
For the relations before and after, the value of the temporal distance is given as a 
duration, using an expression of this form: 
duration (N, Type) 
The value of the duration is obtained by calculating the period corresponding to 
N periods of type Type. For example, the adverbial deuxjours apr~s le ddpart de Paul 'two 
days after Paul's departure' would be represented as (023 represents the occurrence of 
Paul's departure): 
afler(\[hoc, -, -\], \[O23, -, -\], duration(2, day)) 
If the temporal distance is not known (or irrelevent), it is indicated by indefinite. 
Thus, aprbs le ddpart de Paul 'after Paul's departure' would be represented as: 
after(\[ttoc, -,-\], \[023,-, -\], indefinite) 
103 
Computational Linguistics Volume 22, Number 1 
Now, let's suppose that the localizer ti is of type T. tn some cases, a good way 
to express it is by giving its position relative to another time tj of the same type. For 
example, the ATL cinq jours plus tard 'five days later' is not used to mean "at a time 
in the future, five days from the reference time," but rather "the 5th day after the one 
which included the reference time." If the reference time is itself a day, the semantics 
of this ATL could be: 
relpos( 5, \[hoc, day, _\], \[trey,-,-\]) 
If trey is not a day, we must express the relation by taking as anchor the day which 
includes it: 
relpos (5, \[ Gc, day, _\], incl ( \[ t l, day, _\], \[trey, day, -\])) 
(This takes for granted that the time trey is not bigger than a day. If trey were bigger 
than a day, it would not make any sense to express relative position by specifying the 
temporal distance in days.) 
Similarly, hier 'yesterday' would be expressed semantically as "the day that is the 
first one before the day that includes the time of speech": 
relpos(-1, \[tloc, day, -\], incl ( \[\[h, day, _\], \[n, _, _\])) 
We have seen a way of expressing duration, by giving the length as a number 
of time units. But there is another way of expressing duration: by indicating the two 
boundaries of the period. By using this method, not only the duration of an occurrence 
is expressed, but so is its temporal location, at least partially. The relation extent is used 
to express this kind of duration. For example, the semantics of du 3 avril au 5 mai 'from 
April 3rd to May 5th' is formulated as: 
extent(\[hoe, _, _\], inclin ( \[ h, day, 3\], \[t2, month, april\]), inclin ( \[t3, day, 5\], \[t4, month, may\])) 
The semantics of du 3 au 10 avril 'from April 3 to 10' should represent the fact that 
the whole duration is included in the same month: 
inclin (extent (\[boo -, -\], \[h, daY, 3\], \[t2, day, 10\]), \[t3, month, april\]) 
The relation extent is also used to represent adverbials like depuis trois jours 'for 
three days' and pendant trois jours ?z partir du 3 avril 'during three days starting from 
April 3rd'. These adverbials explicitly express one of the two boundaries. In the first 
example, it is either the time of speech or the reference time (the ATL means "for 
three days until now" or "for three days until then"). In the second example, it is the 
time corresponding to April 3rd. The other boundary is indicated implicitly by giving 
a temporal distance from the anchor. The first example, supposing that the explicit 
boundary is speech time, would be represented as: 
extent(\[tloc, -, -\], before(It1, -, -\], \[n, _, _\], duration (3, day)), In, _, _\]) 
This expresses a period whose ending is the time of speech and whose beginning must 
be calculated by finding the time that is three days before speech time. 
The second example would be represented as: 
extent( \[tloc, _, _\], inclin (\[tl, day, 3\], \[t2, month, april\]), after(It3, -, -\], \[t l, -, -\], duration(3, day))) 
Note that in both examples, the same temporal constant represents both the explicit 
boundary and the anchor of the relation after or before used to express the implicit 
boundary (n and tl, respectively). 
104 
Gagnon and Lapalrne From Conceptual Time to Linguistic Time 
Table 2 
Relative localizers. 
Adverbial Semantics 
entre le 3 avril et le 10 mai inclin(\[hoo_,d,extent(\[ti,_,d, 
(between April 3rd and May lOth) inclin(\[t2,day,3\],\[t3,month,april\]), 
inclin(lt4,day,10\],lts,month,aprill))) 
jusqu'tt il y a trois jours end(\[h ..... -\], 
(until three days ago) before(lt2,_,_\],\[n,_,d,duration(3,day))) 
jusqu'/i mercredi de cette semaine end(\[tloo_,_\],inclin(\[t2,moment-of-day,wednesday\] (until Wednesday of this week) incl(\[tg,week,_\],\[n,_,_\]))) 
Considering the examples we have just given, one may think that recursivity 
applies only to the anchor. This is not the case. The triplet that gives the location time 
in the expression can be replaced by a complex expression. We have such a situation 
with le 3 avril dernier 'the last April 3rd' represented as: 
relpos (- 1, inclin ( \[tlo¢, day, 3\], \[tl, month, april\]), incl( \[t2, day, _\], \[n, _, _\])) 
More precisely, this expression means "the April 3rd that is the first one in the 
past, taking speech time as starting point." 
Finally, to illustrate the richness of our semantics for expressing ATLs, in Table 2 
we give a list of more adverbials with their semantics. Note the extensive use of the 
combination property. 
Thus, to specify the localization of an occurrence, we can use another simultaneous 
object, or use a localizer that is expressed in relation to another localizer. The list of 
relations given in Table 1, together with the possibility of combining them, offers a very 
powerful way of expressing the great diversity in the semantics of temporal localizers. 
Of course, not all the combinations may be expressed naturally in the language, but 
we are convinced that most of the ATLs can be expressed with this semantics. 3 
The problem of representing temporal location has received a lot of attention in 
the past (Dowty 1979, 1982, 1986; Bach 1986; Verkuyl 1989; and Vlach 1993). But these 
works have focused on the aspectual structure of adverbials and their relation to tense 
and aspect. We have not found any previous proposals of a recursive semantics like 
ours for representing the various types of localizations. 
More related to our work is Allen (1983) who proposes a set of primitive relations 
to represent all possible relations between two temporal intervals. The relations defined 
in Table I differ in many respects from the relations proposed by Allen. As mentioned 
before, ours are less precise. For example, the moment represented using the relation 
end in our model corresponds to three relations in Allen's model. Suppose that t* and 
tj* are the objective times corresponding to the localizer and the anchor, respectively. 
Then, in Allen's model, the end of t~ could coincide with the beginning of tj*, could 
coincide with the end of tj*, or could be included in t 7. 
The main reason for using a different set of primitives is to represent as closely 
as possible the way temporal localization is dealt with in language. The result of our 
3 In fact, the set of relations described here is not sufficient. In Gagnon and Bras (1994), we define a more complete set of relations. 
105 

Gagnon and Lapalme From Conceptual Time to Linguistic Time 
I 
t ENVIRONMENT 
...... ------o-M"-'--"~-~'- ........ n- - -Knowte.l.e.,; I base 
I structure concepts t 
I 
I 
I 
I 
I 
I 
! 
! 
I 
! 
_J 
P 
R 
E 
T 
E 
R 
I SEMANTIC 
INTERFACE 
I 
-.~ ENGINE < 
¢ 
GRAMMAR 
t (Systemic 
I t network) 
I 
I 
Ii I t I t < --.... -*BLACK 
E I BOARD 
I R 
I 
Figure 8 
Implementation of the surface generation process in Prdtexte. 
To produce a sentence, the network is processed from left to right. When a system 
is entered, a choice is made that may lead to another system or to a conjunction of 
systems processed concurrently. The syntactic structure of the phrase results from a 
set of constraints determined by features selected during the traversal of the entire 
network. 
The choices made in the first traversal of the network determine the overall struc- 
ture of the sentence, represented as an ordered sequence of functions that must be 
fulfilled. The term "function," in this context, refers to the role played by a con- 
stituent of a phrase in achieving a communicative goal (Halliday 1985). For example, 
a sentence can often be decomposed into three constituents fulfilling the following 
functions: Subject, Predicate, and Object. 4 
Once the functional structure of the sentence is established, its network is traversed 
again to determine the syntactic structure, which is then further refined until each 
function is realized by a single word. 
Figure 8 illustrates the organization of the modules in Pr6texte, inspired by Nigel 
(Mann 1983; Mann 1985; Matthiessen 1985; Matthiessen and Bateman 1991). To produce 
a sentence, Pr6texte uses three information components: the environment, containing 
the information about the message and a knowledge base describing how to achieve 
lexicalization; the grammar, represented as a systemic network; and the blackboard, 
used to determine the syntactic structure. The engine controls the surface genera- 
tion process and accesses the three information components through three interface 
modules: semantic interface, interpreter and realizer. The solver determines the fi- 
nal structure of the constituent, using constraints posted in the blackboard during the 
traversal of the network. 
4 In this text, names of functions will always be capitalized. 
107 
Computational Linguistics Volume 22, Number 1 
Before starting the surface generation process, the environment is augmented with 
information that determines the message: 
• a semantic structure such as the one illustrated in Figure 4; 
• a set of relevant concepts, which are the elements of the conceptual 
representation pertaining to the entities in the semantic structure; 
• some pragmatic information, which specifies how to transmit the 
message. 
The engine starts by posting on the blackboard the description of the realized con- 
stituent representing the sentence. It then activates the traversal of the network by the 
interpreter. To select a feature in a system, the interpreter transmits inquiries to the 
engine. If the information necessary to respond to the inquiry is in the environment, 
the engine transmits the inquiry to the semantic interface. If an inquiry is about a 
decision previously made in the surface generation process (for example, a question 
about what features have been selected in a system visited earlier), then the engine 
transmits the inquiry to the realizer. 
Answers to inquiries enable the selection of a feature in the visited system. The 
interpreter then extracts a set of realization statements associated with the selected 
feature. After the execution of these statements by the realizer, the information about 
the structure of the realized constituent, contained in the blackboard, is updated. Three 
kinds of action may be executed by the realizer: 
• addition of a new constituent with the appropriate semantic information 
fulfilling a specific function; 
• updating of the information pertaining to one constituent; 
• addition of partial ordering constraints that identify the sequence of 
functions composing the final structure. 
This process goes on until no more system can be visited. The solver then solves 
the ordering constraints on the blackboard. We thus obtain a sequence of functions 
that constitutes the final structure of the realized constituent. Each of these functions 
is associated with semantic information extracted from the environment. For example, 
the sentence may contain the function Temp_loc (temporal localizer), whose semantic 
information will be the expression associated with the temporal marker LOC in the 
input. If a function is to be lexicalized as a word, the lexicon is consulted to identify 
the word, taking into account the features selected during the traversal. Otherwise the 
grammar is re-entered using the function as the new realized constituent with some 
features preselected. For example, to generate the sentence in 2: 
(2) Jusqu'a 10h50, il a survol6 l'Alg6rie 'until 10:50, it flew over Algeria' 
the following semantic structure and relations are posted in the environment: 
message = event(olo) 
N=n 
R =09 
PERSP = n 
LOC = end(\[hl, _, _\], inclin(\[t6, minute, 50\], \[t4, hour, lO\])) 
(a) o10 < n 
(b) o10 C) tu 
108" 
Gagnon and Lapalme From Conceptual Time to Linguistic Time 
Positioner \[ 
Jusque I 
A A /\ 
Reference Zon 4 il \ 
I 1 
h lOh50 
a survol4 l'Alg6rie 
Figure 9 
The structure of jusqu'd 10h50, il a survold l'Alg&ie. 
where o10 must be expressed as an event and tll is a period terminating at 10:50. The 
event o10 is before the time of speech and it coincides more or less with the time used 
as localizer. The first traversal of the grammar determines the overall structure of the 
sentence made of the four functions at the top of Figure 9. As none of them may be 
lexicalized directly as a word, they must be realized by re-entering the grammar. 
We describe only the realization of the function Tempdoc. The semantics associ- 
ated with Temp_loc is the value of LOC in the semantic representation. Traversal of 
the network for the realization of Temploc results in a structure with two functions: 
the Positioner and the Reference Zone. The first function is lexicalized with the prepo- 
sition jusque; for the second function, the grammar is re-entered, taking as semantics 
the anchor of the expression associated with Tempdoc: inclin(\[t6,minute,50\], \[t4,hour, lO\]). 
Again, this results in a structure with two functions; in this case, both functions can 
be lexicalized, ending the realization of the function Tempdoc. The same kind of pro- 
cessing is done for the other functions in the sentence. 
5. The Production of ATLs 
In Section 3.3, we showed how the semantics of ATLs is represented; in this section, we 
present how ATLs can be lexicalized. Table 3 gives a list of semantic representations 
and their translations into ATLs produced by our generator. As in Section 3.4, we use 
ttoc for the temporal constant corresponding to the localizer of the occurrence, n for 
speech time and tref for the reference time. 
5.1 Syntactic Compositions 
Some ATLs (1-7, 16) are simple, while others (8-15, 17-18) contain an embedded tem- 
poral adverbial. For example, in (11) jusqu'?l mercredi de cette semaine 'until Wednesday 
of this week' contains another ATL, which itself contains another embedded ATL cette 
semaine 'this week'. 
109 
Computational Linguistics Volume 22, Number 1 
Table 3 
List of adverbial temporal locations. 
Semantics ATL 
(1) relpos(-1, \[tloc, day, _\], incl(\[t2, day, _\], \[n, _, _\])) 
(2) relpos(-1, \[hoc, daY,-\], incl(\[t2, day, _\], \[tref, _, _\])) 
(3) incl(\[tloc, day, _\], \[n, _, -\]) 
(4) incl(\[tloc, month, _\], \[n, _, _\]) 
(5) incl(\[tloc, month, _\], \[tref, -, -\]) 
(6) \[tloc,month,April\] 
(7) \[01 ,-,-\] 
(8) inclin(\[tloc,moment-of-day, morning\], 
inclin(\[t2,season,summer\],\[t3,year,1995\])) 
(9) inclin(\[ tloc,half-hour,1\],occurrence(ol ) ) 
(10) duration(3,day) 
(11) end(\[tloc,_,_\],inclin(\[t2,day, wednesday\], 
incl(\[t3,week,_\],\[n,_,_\]))) 
(12) begin(\[tloc,_,_\],relpos(1,\[t2,month,_\], 
incl (\[ t 3,month,-\],\[ tref ,-,-\]) ) ) 
(13) begin(\[tloc,_,_\],inclin(\[t2,day, lO\],\[t3,month,may\])) 
(14) relpos(3,\[tloc,day,_\],occurrence(ol )) 
(15) after(\[tloc,day,_\],\[tref,_,_\],duration(3,day)) 
(16) after(\[tloc,_,_\],\[tref ,_,_\],indefinite) 
(17) extent(\[tloc,_,_\], 
inclin(\[t2,day,3\],\[t3,month,april\]), 
inclin(\[t4,day, lO\],\[ts,month,may\])) 
(18) extent(\[tloc,_,_\], 
before(\[t2,-,-\],\[tre f ,-,-\],duration(3,day)), 
\[tre/,-,-l) 
hier (yesterday) 
la veille (the day before) 
aujourd'hui (today) 
ce mois-ci (this month) 
ce mois-la (that month) 
en avril (in April) 
quand Robert est parti (when Robert left) 
le matin du 3 avril 1995 (the morning of April 3rd 1995) 
la premi6re demi-heure de l'6mission (the 
first half hour of the program) 
durant trois jours (during three days) 
jusqu'a mercredi de cette semaine (until 
Wednesday of this week) 
a partir du mois suivant (from the following 
month) 
depuis le 10 mai (sinceMay lOth) 
trois jours apr6s le d6part de Robert (three 
days after Robert's departure) 
trois jours plus tard (three days later) 
puis (then) 
du 3 avril au 10 mai (from April 3rd to May 
lOth) 
depuis trois jours (since three days) 
Unfortunately, the combination of localizers in the semantics does not always 
correspond to the combination of adverbials. For example, if there were such a corre- 
spondence, the adverbial in (1) would be something like lejour avant lejour qui contient 
l'instant d'~nonciation 'the day before the day that contains the time of speech'. Instead, 
we get the simple adverbial hier 'yesterday'. For complex semantic expressions, in all 
examples except (15), (16) and (18), there is one embedded adverbial corresponding to 
each anchor. For example, the anchor relpos(1,\[t2,month,_\], incl(\[tB,month,_\], \[tref,-,-\])) in 
(12) corresponds to le mois suivant 'the following month' in the adverbial. 5 Examples 
(1-5) are special, since the relation and the anchor are combined in the same syntactic 
structure. In (15), direct translation of the anchor into an adverbial would produce 
trois jours apr~s ce mornent-l?l 'three days after that moment', and in (18), pendant trois 
jours jusqu'?~ ce moment-l?l 'during three days until that moment'. 
Since there is not always a direct correspondence between semantic and syntactic 
5 Note that in the adverbial, du is a contraction of de le. 
110 
Gagnon and Lapalme From Conceptual Time to Linguistic Time 
QTY" ANCHORS 
I unique 
double ( 
Figure 10 
Section of Pr4texte's grammar for adverbials. 
ANCHOR 
ANCHORI 
ANCIIOR2 
deictic 
~ anaphoric 
autonomous 
~ - deictic anaphoric 
autonomous 
i deictic 
anaphoric 
autonomous 
forms, which one should be used in the grammar to distinguish among ATLs? We 
have chosen the semantic form because adverbials are distinguished not only by the 
number of anchors but also by their nature. Examples (4) and (5) are both syntactically 
simple---the anchor in the semantic form is not expressed--but it is the anchor that 
explains their difference: the first uses speech time, whereas the second uses reference 
time. 
Figure 10 illustrates the part of the network taking into account the combination 
property. We first identify the number of anchors. If there is only one, the feature unique 
is selected in the system QTY_ANCHORS. Otherwise, double is selected. Then, for each 
anchor, we must establish if it is deictic, anaphoric, or autonomous. If the localization 
represented by the anchor is made in relation to the time of speech, deictic is selected; 
if the localization is made in relation to the reference time, anaphoric is selected; if 
the anchor achieves a localization without using either of the two temporal markers, 
autonomous is selected. In Table 4, we indicate the features selected in the systems of 
Figure 10, for the production of the adverbials given in Table 3. 
Some adverbials may be distinguished using the systems of Figure 10, but Table 4 
shows that these systems are not enough. The features have a strong influence on the 
most embedded adverbials. For example, in (11), the selection of deictic results in cette 
semaine 'this week' for the most embedded adverbial but if the feature anaphoric had 
been selected, it would have produced cette semaine-lit 'that week'. 
These features alone do not explain the recursive form of adverbials. Figure 11 
shows the structure of two adverbials from Table 3. The structure of adverbial (11) is 
given in (a). It has three levels, each one corresponding to one adverbial. The simple 
structure of adverbial (4) is shown in (b). Their difference is not only due to the number 
of levels in the structure. In (a), the structure contains a function, the Positioner, that 
expresses the relation to the anchor; there is no function in (b). Sometimes an anchor 
is not realized syntactically at all. In Figure 12, we consider (13) and (15): the anchor 
is expressed in (13), shown in (a), but not in (15), shown in (b). The structure of (15) 
contains the Positioner and a function conveying the temporal distance to an implicit 
anchor. 
These examples show that features are not sufficient; to determine the syntactic 
structure of the adverbial, we need more systems in our grammar, such as the network 
of Figure 7. The two networks of Figure 7 and Figure 10 must be traversed in parallel. 
The system ZONE_DESIGNATION first distinguishes between adverbials that ex- 
111 
Computational Linguistics Volume 22, Number 1 
Table 4 
Distinction of adverbials using the anchor. (The numbers correspond to the adverbial's 
position in Table 3.) 
Adverbial QTY_ANCHORS ANCHOR ANCHOR1 ANCHOR2 
(1) hier 
(3) aujourd'hui 
(4) ce mois-ci 
(11) jusqu'h mercredi de cette semaine 
(2) la veille 
(5) ce mois-lh 
(12) a partir du mois suivant 
(15) trois jours plus tard 
(16) puis 
(6) en avril 
(7) quand Robert est parti 
(8) le matin du 3 avril 1995 
(9) la premi6re demi-heure de l'6mission 
(10) durant trois jours 
(13) depuis le 10 rnai 
(14) trois jours apr6s le d6part de Robert 
(17) du 3 avril au 10 mai 
(18) depuis trois jours 
unique 
umque 
unique 
umque 
umque 
unique 
unique 
unique 
umque 
umque 
unique 
unique 
umque 
unique 
umque 
umque 
double 
double 
deictic 
deictic 
deictic 
deictic 
anaphoric 
anaphonc 
anaphoric 
anaphoric 
anaphoric 
autonomous 
autonomous 
autonomous 
autonomous 
autonomous 
autonomous 
autonomous 
autonomous 
anaphonc 
autonomous 
anaphoric 
Positioner \[ 
Jusque 
Until 
Reference Zon~ \ 
\[Zone Designator\[ Attributor Reference Zon~ 
h mercredi de \[,-Zone Designator\] 
Wednesday of / NNN, N 
cette semaine this week 
(a) (b) 
Figure 11 
Difference of structure for (11) and (4). 
Zone Designator\] /\ 
ce mois-ci this month 
press localization directly, as in (1-9), and adverbials that relate it to other localizers. 
Selection of direct includes a function Zone Designator in the structure for the most 
embedded adverbial of (a) and the adverbial of (b) in Figure 11. This function is real- 
ized by a phrase expressing the temporal location, which may be a temporal constant 
(feature chronological) or an occurrence (feature occurrential). 
If relational is selected in ZONE_DESIGNATION, a function Positioner is inserted 
in the structure. This function is realized by a phrase that expresses the relation of the 
localizer to its anchor. There are two types of relational localizers: those that express a 
112 
Gagnon and Lapalme From Conceptual Time to Linguistic Time 
Positioner \[ RefercnceZonc( /\ 
depuis le 10 mai 
Since May lOth 
(a) 
Figure 12 
Difference of structure for (13) and (15). 
I T m ora,  sta eol pvs,tionor I 
trois jours plus tard 
three days later 
(b) 
Table 5 
Distinction of adverbials with type of designation. (The numbers correspond to Table 3.) 
Adverbial ZONE DESIGNATION TYPE LOC ZONE LOCT 
ASPECT 
(1) hier direct chronological -- 
(2) la veille direct chronological -- 
(3) aujourd'hui direct chronological -- 
(4) ce mois-ci direct chronological -- 
(5) ce mois-la direct chronological -- 
(6) en avril direct chronological -- 
(8) le matin du 3 avril 1995 direct chronological -- 
(9) la premi6re demi-heure de l'6mission direct chronological -- 
(7) quand Robert est parti direct occurrential -- 
(10) durant trois jours relational -- durative 
(11) jusqu'~ mercredi de cette semaine relational -- durative 
(12) a partir du mois suivant relational -- durative 
(13) depuis le 10 mai relational -- durative 
(17) du 3 avril au 10 mai relational -- durative 
(18) depuis trois jours relational -- durative 
(14) trois jours apr6s le d6part de Robert relational -- punctual 
(15) trois jours plus tard relational -- punctual 
(16) puis relational -- punctual 
duration (see (a) in Figure 12) and those that designate a punctual temporal location 
(see (a) in Figure 11 and (b) in Figure 12). 
The classification of adverbials using these distinctions is shown in Table 5. But 
even by combining this classification with that of Table 4, we cannot distinguish be- 
tween all adverbials. For example, (1, 3, and 4) select the same features in both net- 
works, as do (2 and 5), (6, 8, and 9), and (15 and 16). For each of the four cases of 
Table 5, we will show how the adverbials can be distinguished. 
5.2 Relational Localizers 
5.2.1 Punctual Localizers. The function Positioner is always present in adverbials for 
which the feature punctual has been selected; this is a consequence of the selection of 
the feature relational. In addition to Positioner, there can be two more functions. One 
is the Temporal Reference Zone, which conveys the localizer to which the relation ex- 
pressed by the Positioner pertains. In our list, only adverbial (14) contains this function: 
le ddpart de Robert 'Robert's departure'. The other function is Temporal Distance, which 
expresses the length of time from the localizer used as anchor. This function occurs in 
113 
Computational Linguistics Volume 22, Number 1 
Positioner l 
Puis Then 
Auparavant 
Before 
(a) 
\[ Positioner \] Temporal Distance \[ 
Dans trois jours 
In three days from now 
I1 y a trois jours Three days ago 
(b) 
\[ T~Zone\[ Positioner \] 
Trois jours pllas tard Three days tater 
Trois jours av ant Three -days earlier 
(c) 
\[ Positioner \] Temp Ref Zone \[ 
a~re~s le d61~art de Robert R o berY s departure 
avant 8h00 before 8:00 
\] Temporal Distance \] Positioner \[ Temp Ref Zone\[ /\ /\ 
T,~trOiS jours al~r~s !e d6partde Robert ee aays after Kooert's aeparture 
Trois jours avant le d6part de Robert Three ~ays before Robert's departure 
(d) (e) 
Figure 13 
Structure for punctual adverbials. 
RELATION TYPE V before punctual 
L_ after 
f explicit REFERENCE ZONE E implicit 3-- 
TEMtK)RALDISTANCE F definite 
~-- indefinite 
Figure 14 
Grammar section for relational punctual adverbials. 
adverbials (14) and (15): trois jours 'three days'. Thus, (14) contains both functions and 
(15) contains only the Temporal Distance. Adverbial (16) is distinguished from (14) 
and (15) because it contains neither of these functions. 
In dans trois jours 'in three days from now' and apr~s le 9 octobre 'after October 9th', 
we find two different structures. The elements of the first structure are exactly the 
same as the structure of adverbial (15), but they occur in a different order: Temporal 
Distance comes before the Positioner. In the second structure, there is a Temporal 
Reference Zone, le 9 octobre, but no Temporal Distance. Thus, for punctual adverbials, 
there are five possible structures, illustrated in Figure 12. To distinguish between these 
adverbials, we use a network, part of which is shown in Figure 13. 
Two features are expressed by the system RELATION TYPE: before and after. To 
realize the Positioner, there is no need to re-enter the grammar, since it may be found 
directly in the lexicon. The lexical choice depends not only on the selection achieved 
114 
Gagnon and Lapalme From Conceptual Time to Linguistic Time 
Table 6 
List of punctual adverbials. 
Adverbial RELATION REFERENCE TEMPORAL 
TYPE ZONE DISTANCE 
(14) trois jours apr6s le d6part de Robert after explicit definite (three days after Robert's departure) 
apr6s le 9 octobre (after October 9th) after explicit indefinite 
(15) trois jours plus tard (three days later) after implicit definite 
dans trois jours (three days from now) after implicit definite 
(16) puis (then) after implicit indefinite 
trois jours avant le d6part de Robert before explicit definite (three days before Robert's departure) 
avant 8h00 (before 8:00) before explicit indefinite 
trois jours avant (three days earlier) before implicit definite 
il y a trois jours(three days ago) before implicit definite 
auparavant (before) before implicit indefinite 
in RELATION TYPE, but also in the choice made in the system ANCHOR of figure 10. 
For example, in the cases where after is chosen, the Positioner could be lexicalized as 
puis 'then' or plus tard 'later'; if anaphoric is chosen in ANCHOR it can be lexicalized 
as dans 'in' if deictic is chosen or apr~s 'after' if autonomous is chosen. 
The fact that Temporal Distance and Temporal Reference Zone are optional in the 
structure is represented in the grammar by two parallel systems: REFERENCE ZONE 
and TEMPORAL DISTANCE. If, in REFERENCE ZONE, explicit is chosen, the function 
Temporal Reference Zone is included in the structure. Since this function represents 
another localizer, the anchor, it is realized by re-entering the grammar, taking as input 
the semantic representation of this anchor. In TEMPORAL DISTANCE, the selection 
of definite results in the inclusion of the function Temporal Distance. To realize it, 
the grammar must be re-entered, and some features must be preselected so that it is 
realized as a noun phrase. 
Table 6 lists all possible adverbials represented by the network of Figure 13 to- 
gether with their selected features. The three adverbials taken from Table 3 are pre- 
ceded by their reference number to ease the comparison of their semantics with the 
selected features; their relations will be discussed later. 
The distinction between structures (b) and (c) in Figure 12 is not explained by 
the grammar section shown in Figure 13, since the same features are selected for trois 
jours plus tard and dans trois jours. However, structure (b) in Figure 12 is found only 
for deictic localizers. Therefore, features for deictic localizers selected in the system 
ANCHOR, will distinguish structure (b) from structure (c). 
Let us now see how the features are selected for the production of relational 
punctual localizers. In RELATION TYPE, the feature reflects the relation used in the 
semantics, if this relation is before or after. Adverbial (14) deserves some explanation. 
Since its semantic expression uses the relation relpos, we would expect its syntactic 
realization to be: le troisi~me jour apr~s le jour du d6part de Robert 'the third day after the 
day of Robert's departure', but this usage is rare. Instead we find trois jours apr~s le 
d6part de Robert 'three days after Robert's departure', which is what we would expect if 
the semantic expression used the relation after. This seems to be because if the temporal 
distance is one unit, a direct localizer is preferred. So, instead of generating un mois 
plus tard 'one month later', we produce le mois suivant 'the next month'. Our intuition 
115 
Computational Linguistics Volume 22, Number 1 
durative 
DURATION 
TYPE 
bound 
quantified 
DURATION 
ANCHOR 
DURATION 
PERSPECTWE -- anterior 
posterior 
double 
QUANTDURATION E anterior 
ANCHOR posterior 
nil 
Figure 15 
Grammar section for relational durative adverbials. 
-- internal 
external 
\[P.~i--, \] D~oo Q~,ty i l I '°'ia°'= \] Bo.d~, \] 
..~e~ ~ays . Jusque Is mercredi de 
three days Sm.ce 
for ~¢rom that moment) A r~oir de le 10 mai 
may lOth 
(a) (b) 
Figure 16 
Structure for durative localizers. 
Du 3 avril au lOmai 
From April 3rd to May 10th 
(c) 
is that when X is '%ig" we have this equivalence: 
relpos(X, \[ti, Ti, Ni\], Z) ::~ afler(\[ti, Ti, Ni\], Z, duration(X, Zi) ) 
and similarly for X negative and the relation before. More study is needed to determine 
the threshold at which the two relations become equivalent in the linguistic realization. 
We are sure that for X = I or X = -1, they are not equivalent, so in our implementation, 
we use 2 and -2 as thresholds. 
In the system REFERENCE ZONE, the feature implicit is chosen if the anchor is 
a simple localizer using the reference time or the time of speech, otherwise explicit is 
chosen. Feature selection in TEMPORAL DISTANCE depends on whether the third 
argument in the semantic expression is indefinite or a specified duration. 
5.2.2 Durative Adverbials. We now show how the durative adverbials of Table 5 can 
be differentiated. The part of the network that generates these adverbials is shown 
in Figure 14. We give the three kinds of structure identified for these adverbials in 
Figure 15, and, finally, in Table 7 we list the durative adverbials of Table 5 with their 
corresponding features according to the systems of Figure 14. To give a complete 
illustration of all adverbials generated with the network of Figure 14, we added one 
adverbial to the list: pendant trois jours 'for three days from now', which is symmetrical 
to depuis trois jours 'for three days until now'. 
In Figure 14, we distinguish two types of durative adverbial phrases: bound, if 
the duration is expressed by specifying one or two of its boundaries; and quantified, 
116 
Gagnon and Lapalme From Conceptual Time to Linguistic Time 
Table 7 
List of durative adverbials. 
Adverbial DUR. DUR. DUR. QUANZ 
TYPE ANCH. PERSE DUR. 
ANCH. 
(13) depuis le 10 mai (since May lOth) bound ant. int. -- 
(12) a partir du rnois suivant bound ant. ext. -- 
(from next month) 
(11) jusqu'~ mercredi de cette semaine bound post. -- -- 
(until Wednesday of this week) 
(17) du 3 avril au 10 rnai bound double -- -- (from April 3rd to May lOth) 
-- pendant trois jours quant. -- -- ant. 
(for three days until that moment) 
(18) depuis trois jours quant. -- -- post. 
(for three days from that moment) 
(10) durant trois jours (during three days) quant. -- -- nil 
if the duration is expressed as a quantity of time units. In the first case, we get a 
structure such as (b) or (c) in Figure 15. In (b) there is only one boundary, and the 
Positioner indicates which one is used: its lexicalization depends on the feature selected 
in DURATION ANCHOR. If anterior is selected, one further distinction is required to 
lexicalize the Positioner, as represented by the system DURATION PERSPECTIVE. The 
Positioner is realized by the phrase ~ partir de if the feature chosen is external, otherwise 
it is realized as depuis. The choice depends on the aspect of the occurrence reported: 
the feature external is chosen if the occurrence is presented as an event, and internal is 
chosen if the occurrence is presented as a situation. These two cases are exemplified 
in the following two sentences: 
(3) a. A partir de 1972, il enseigna a l'Universit6 de Montr6al. 'From 1972 
on, he taught at Universit6 de Montr6al'. 
b. Depuis 1972, il enseignait a l'Universit6 de Montr6al. 'Since 1972 he 
was teaching at Universit6 de Montr6al'. 
In (3a), since the occurrence is presented as an event, the feature external is selected 
during the determination of the ATL, thus resulting in the form ~ partir de 1972 'from 
1972'. In (3b), the same occurrence is presented as a situation considered from the 
reference time. Thus, internal is selected, resulting in the form depuis 1972 'since 1972'. 
This is a good example of the interaction of ATLs with the aspect of the occurrence. 
In the interpretation of the ATL in (3b), the duration is anchored not only on 
the year 1972 but also on the reference time included in the occurrence. But in the 
semantics of the ATL, which uses the relation begin, as well as ATL (12) in Table 3, 
there is only one anchor. Even if the reference time is involved in the understanding 
of the whole sentence, it is not directly expressed in the semantics of the ATL. 
An alternative would be to express the same localization using the relation extent, 
as in (18) of Table 3. If the reference time is included in 1982 (the beginning of the 
duration expressed in (3b) thus being 10 years before), the semantics would be: 
extent(\[hoc,_,_\],before(\[tl,year, d,\[t2,_,-\],duration(lO,year)), 
incl(\[ta,year, d,\[tref ,_,_\])) 
117 
Computational Linguistics Volume 22, Number 1 
In depuis dix ans 'For 10 years', the meaning of the ATL, which is "since 10 years 
in the past starting from this moment" requires the use of the reference time. 
If, in the system DURATION ANCHOR, the feature double is chosen, we get a 
structure containing two boundaries, as in (c) in Figure 15. A boundary, in the structure 
of a bound localizer, is always realized as a temporal adverbial, by re-entering the 
grammar. 
When the feature quantified is selected, the structure in (a) of Figure 15 is obtained. 
To realize the Positioner in this case, another system is required, because the quantity 
of time that constitutes the duration can be worded in many ways. We can express 
the duration of the occurrence without giving any hint about its location in time, as 
in durant trois jours 'during three days', or we can indicate a duration that begins or 
ends at some time. 
To see how the features are selected in the grammar section of Figure 14, compare 
the adverbials of Table 7 with their semantics as given in Table 3 (the semantics for 
pendant trois jours is the same as depuis trois jours, but the relation after is substituted 
before and the two anchors are reversed). First, the feature bound in DURATION TYPE 
is selected if the relation used in the semantic representation is either begin or end, or 
if it is extent and the two anchors are autonomous. In DURATION ANCHOR, features 
corresponding to these three cases are selected. In DURATION PERSPECTIVE, the 
selection depends on the aspect of the occurrence reported in the sentence. 
In DURATION TYPE, quantified is selected if the semantics uses the relation extent 
and one anchor is deictic or anaphoric, as in our examples, or if it uses the relation 
duration. In the first case, the selection depends on the position of the anaphoric or 
deictic anchor in the expression. If the relation duration is used, since there is no anchor, 
the feature nil is selected in the next system. 
5.3 Direct Localizers 
We complete our discussion of adverbials by explaining how the direct localizers can be 
differentiated. Figure 16 shows the structure of the direct adverbials that constitute the 
first half of Table 5; there are three possible structures for a direct localizer. The simplest 
ones, in (a) and (b), contain only one function, Zone Designator, that expresses the 
temporal location zone designated by the adverbial. This function is realized directly 
by an adverb, in (a), using the lexicon depending on the system ANCHOR of Figure 10. 
In (b), the grammar must be re-entered to generate a nominal phrase whose form also 
depends on the choice in ANCHOR. 
In some cases, it is not sufficient to specify a temporal location zone: we must also 
add what we call a Pointer to relate the occurrence with this zone. In our examples, 
the Pointer indicates that the occurrence takes place during the month of April, or when 
Robert left. The existence of such a localizer in the structure seems to depend on the 
level of the adverbial in the embedding structure. For example, we find a Pointer in 
the adverbial ~ ShOO 'at 8:00' if it is used alone, but not if it is embedded in another 
adverbial, like depuis 8hO0 'since 8:00'. Our approach to this problem may be contrasted 
with Forster's (1989), who determines the realization of the Pointer by the temporal 
aspect of the Zone Designator (durative or punctual) using a constraint propagation 
technique. 
Other possible structures for direct localizers are illustrated in (d) and (e) in Fig- 
ure 16. One function is the Zone Designator, which designates the direct expression 
of temporal zone. If this zone is included in another localizer or in a position relative 
to another localizer, we must include another function in the structure: the Reference 
Zone, which corresponds to this second localizer. The Attributor links these two func- 
tions. In (e), le matin du 3 avril 1995 'the morning of April 3rd 1995' directly expresses 
118 
Gagnon and Lapalme From Conceptual Time to Linguistic Time 
Izo=~i~o~l Iz/;n~sign~,rl \[Po~i~ tz°nem~'a~"°q /X /X 
Aujourd'hui Ce mois-ci En avril Today This month In April 
l-Iier Ce mois-l~t Quand Robert est parti Yesterday That month When Robert lift 
La veille 
The day before 
(a) Co) (c) 
I ZoneDesignat~rAttributor \] ReferenceZonb 
\[Zone Designatot de I Z°ne DesignatOr l /\ 
La premi&e demi-heure 1'6mission 
The first half-hour the program 
(d) 
I Z°neDesignat°1 Attributor Referen¢.eZorle / 
Zone Designatoq Auributor I Kelerence z.o~e /\ I 
d~ | \[ ZoneDestgnator I Attn0utor li~c~crencez-on~ \[ Zone Designator I of I 
the3rd I avril ~ t ator I ! 
Le matin I April ~ I j//N'N I The morning ! I 
1995 I I=_-- I I ............ =_----:-'---- 
(e) 
Figure 17 
Structure for direct localizers. 
a morning. This morning is itself part of another localizer, and so on. The Attributor 
is sometimes lexicalized as an empty item. 
To these direct localizers, we must add the embedded direct adverbials found 
in the relational adverbials of table 5. In Figure 17, we use by dashed-line boxes to 
indicate those that differ from Figure 16. 
The system shown in Figure 18 differentiates among these different forms of direct 
adverbials. For occurrential adverbials, for example, the second adverbial of (b) in 
Figure 16 and the embedded adverbials in (d) in Figure 16 and (b) in Figure 17, the 
occurrence may be nominalized or not. We do not have any satisfying answers to the 
question of how to choose between these two possibilities. We will state only that 
when the adverbial is embedded, a nominalized form may be preferred to another 
embedded adverbial. 
For chronological adverbials, the system AUTONOMOUS ZONE distinguishes 
between those that have an anaphoric or deictic temporal location zone, and those for 
which the temporal location zone is autonomous. Adverbials of the first type always 
have a simple structure: aujourd'hui, hier, demain, ce mois-ci, ce mois-lh, cette semaine, le 
mois suivant. The temporal location zone is different from the anchor. In mercredi de cette 
semaine "Wednesday of this week', the temporal location zone, expressed by mercredi, 
is autonomous whereas the anchor expressed by cette semaine is deictic. When the 
network is traversed the first time, yes is selected in the system AUTONOMOUS 
ZONE. Its only in the second traversal, when cette semaine is generated, that no is 
selected in this system. 
We must further distinguish adverbials with an autonomous temporal location 
zone, by deciding if their structure contains a Reference Zone or not. The feature 
119 
Computational Linguistics Volume 22, Number 1 
I Bouna y I 
Jusque IZoneDesignatorl Attributor I Referencezon4 
Until / 
\[ Pointer \[Zone Designator I /\ 
a mercredi Wednesday 
de of 
! 
r ..... .T ...... 
V \[Zone Designator \] /\ 
cette semaine this week 
........... i 
.................................. ! 
(a) 
\[ Temporal Distance I Positioner \] Temp Ref Zone I /\ 
Trois jours 
Three days 
apr~s after \[Zone Designator \[ /\ 
le d6part de Robert Robert's departure 
I Positionerl Boundary \] 
A partir de \[Zone Designator \[ 
Fro,. /~ 
Depuis 
From le mois prochain 
the next month 
le 10 mai 
May lOth 
........... _1 
(b) (c) 
Figure 18 
Structure for direct localizers. 
TYPE_LOC_ZONE 
AUTONOMOUS no 
I chronological ZONE E yes hNCLUD1NG TIME implicit 
E explicit 
nominalised occurrential 
NOMINAL 
not nominalized 
Figure 19 
Grammar section for direct adverbials. 
implicit, implying the non-existence of Reference Zone, is selected in the system IN- 
CLUDING TIME if the semantic form is a single triplet \[ti Type, Naming\]. 
There is yet another system that decides if there is a Pointer or not, but as the prob- 
lem of the existence of the Pointer is not completely solved and not really important 
to our discussion, we do not consider this system here. 
120 
Gagnon and Lapalme From Conceptual Time to Linguistic Time 
Table 8 
List of durative adverbials. 
Adverbial AUTONOMOUS INCLUDING 
ZONE TIME 
aujourd'hui (today) no 
hier (yesterday) no 
demain (tomorrow) no 
ce mois-ci (this month) no 
ce mois-l& (that month) no 
le mois suivant (the following month) no 
en avril (in April) yes 
mercredi de cette semaine (Wednesday of this week) yes 
cette semaine (this week) no 
la premi6re demi-heure de l'6mission yes 
(the first half-hour of the program) 
le matin du 3 avril 1995 (the morning of April 3rd 1995) yes 
le 3 avril 1995 yes 
avril 1995 yes 
1995 yes 
implicit 
explicit 
explicit 
explicit 
explicit 
explicit 
implicit 
In Table 8, we present the features selected for direct chronological adverbials, 
embedded or not. The systems in Figure 19 do not suffice to distinguish all direct 
adverbials. The selections in these systems must be combined with the selections made 
in the systems of Figure 10. 
5.4 Related Work on the Generation of ATLs 
The problem of temporal localization has already been studied by many researchers, 
but most of them have focused on the aspectual interaction of the adverbials with 
verb tense; the problem of the semantic and syntactic structures of ATLs has been 
neglected. Molin6s' (1990) study from a linguistic perspective characterizes the ad- 
verbials based on noun phrases. Our work extends hers because our computational 
perspective has made us go farther in the formalization. Bras and Molin6s (1993) 
made a similar attempt, but from the perspective of discourse understanding. Since 
the problems of understanding are very different from the problems of generation, we 
could not simply use their method in a "reversed mode." Their method relies on a 
compositional analysis of the language, where all information units extracted from the 
semantic structure are combined to select one meaning for the adverbial. This compo- 
sitional approach is not easily reversible, and it does not provide any insight into the 
selection problem inherent to the generation task. Ehrich (1987) classifies adverbials in 
the context of generation, but she does not cover all the cases presented in this section. 
Concerning the problem of the generation of ATLs, Maybury (1991) shows how the 
notion of focus as used by McKeown (1985) can be extended to include a temporal 
focus that corresponds essentially to the reference point in the Reichenbach model 
(1947). An operation on the temporal focus, in combination with the value of speech 
and event times, selects the temporal adverbial and the verb tense. Since the emphasis 
in this work was on the planning aspect of the task, the variety of adverbials that can 
be generated is limited. 
Forster (1989) explains how the syntactic structure of a temporal adverbial may be 
controlled by semantic information such as the durative or punctual nature of the lo- 
calizer. Essential134 the final structure is obtained by propagating constraints associated 
with each syntactic subpart of the structure. In particular, he focuses on the interaction 
between prepositional phrases and noun phrases. For example, the preposition on is 
121 
Computational Linguistics Volume 22, Number 1 
selected in on Sunday because Sunday is identified as a punctual localizer; this rules out 
in, which implies a durative localizer. We have already presented one problem with 
this approach: it is not clear how the choice of these prepositions can be achieved by 
propagating semantic constraints. The choice of preposition in French is very different 
from English and it often appears to be arbitrary or conventional. Furthermore, many 
aspects of the problem are neglected, such as the type of reference expressed by the 
adverbial: it is not clear how Forster's system can represent the distinctions between 
anaphoric, deictic and autonomous localizers because the link between the semantic 
and syntactic levels is not fully explained. 
Nigel (Matthiessen and Bateman 1991) offers the widest coverage of English but 
the variety of forms for ATLs is quite limited. The temporal localization that may be 
expressed by different types of syntactic structures is represented in Nigel by systems 
dispersed throughout the whole grammar network. For the expression of temporal 
localizers, their grammar is more dependent on the syntactic structure than ours, 
which is mainly determined by the semantics. 
To summarize, our approach departs from previous approaches by covering more 
types of adverbials, by proposing a semantics for localization, and by explaining in 
detail how the different syntactic structures may be obtained from this semantics. 
6. The Production of Verb Phrases 
In our work, we have focused on the generation of adverbials because we felt this 
problem had not received enough attention and because the temporal localization 
achieved by ATLs is more complex and more diversified than that expressed by verb 
tenses. To generate a discourse like Discourse 1, however, we cannot avoid the problem 
of determining the structure of the verb phrase, because part of the localization is 
achieved by the verb, and because of the relations between verbs and adverbials. 
In our implemention of the expression of temporal localization, the relation be- 
tween the verb and the adverbial is taken into account mainly in the deep generation 
process. In the semantic representation, we find traces of this interaction. By keeping 
these decisions in the deep generation process, the verb phrase and the ATL can be 
generated independently. 
Our method for generating the verb phrase takes advantage of the kind of in- 
formation directly represented in DRT: the relation of the occurrence to speech time, 
which we call the primary localization, the aspect of the occurrence, and the presence 
or absence of a perspective point. It is implemented by the grammar section illustrated 
in Figure 19. In Prdtexte, the production of verb phrases requires many traversals of 
the network. First, when the structure of the sentence is determined, choices are made 
regarding localization, aspect, and perspective. After a first traversal of the network, 
the sentence's structure contains a function called Predicate, realized as a verb phrase. 
The grammar must be re-entered to realize the Predicate. The systems visited during 
this second traversal (not shown here) classify verb tenses in French. Most of the se- 
lections during this second traversal were preselected during the first traversal. For 
each verb tense, there is one associated structure, which contains a main verb and one 
or two auxiliaries. To generate each verb or auxiliary, another traversal is needed. 
In the first system of Figure 19, PRIMARY LOC, the selection depends on the 
temporal relation between the localized occurrence and the speech time. The features 
of the systems ASPECT and SIT TYPE reflect the value of aspect in the semantic rep- 
resentation. If the aspect is event, the system PERSPECTIVE determines if this event is 
presented using a perspective. If there is one, another choice must be made regarding 
its type. 
122 
Gagnon and Lapalme From Conceptual Time to Linguistic Time 
,---- past 
/I'~YL~ ~ present 
situation 
ASI~CT 
event ~RS~CTIVE 
E resultirlg 
open 
I. 
Figure 20 
Generation of verb phrase--sentence level. 
perspective 
no perspective 
PERSPECTIVfi a n teri ority 
TYPE 
posteriority 
Table 9 
Production of VP--examples. 
Selections during first traversal Tense selected during Example 
second traversal 
past - situation - resulting 
past - situation - open 
past - event - perspective - anteriority 
past event perspective 
posteriority 
past - event - no perspective 
plus-que-parfait 
pass4 ant6rieur 
imparfait 
imparfait 
conditionnel 
plus-que-parfait 
pass6 compos6 
A 8h00, il avait terrain4. (At 8:00, he had finished.) 
Une fois qu'il eut termin4 (Once he had finished) 
A 8h00, Robert regardait la t616vision. (At 8:00, 
Robert was watching television.) 
J'ai rencontr4 Robert jeudi dernier. I1 partait le 
lendemain. (I met Robert last Thursday. He was going 
to leave the day after.) 
J'ai rencontr6 Robert jeudi dernier. I1 m'a dit qu'il 
partirait le lendemain. (I met Robert last Thursday. 
He told me that he would leave the next day.) 
J'ai rencontr6 Robert jeudi dernier. I1 4tait arriv4 
la veille. (I met Robert last Thursday. He had arrived 
the day before.) 
Robert a parl4 ~ Marie. (Robert talked to Marie). 
Table 9 shows examples of verb phrases, including the list of features selected 
for each example during the two traversals. The same tense can be used for different 
feature patterns. This is the case with the imparfait and the plus-que-parfait: the impar- 
fait expresses an open situation or an anterior perspective, while the plus-que-parfait 
presents a resulting situation or a posterior perspective. This may be a problem in 
an understanding process, since it is a source of ambiguity, but not in a generation 
process since it does not matter if two different inputs map into the same syntactic 
structure. 
More than one verb tense may be used for the same features. This means that 
our grammar is not complete: more systems would be needed to distinguish among 
these different cases. For example, to distinguish the two tenses used with the first 
feature pattern of Table 9, we would have to augment the grammar section of Figure 19 
123 
Computational Linguistics Volume 22, Number 1 
to determine if the verb phrase is part of a temporal adverbial or not. For the two 
cases in the third feature pattern, the difference relates to the use of indirect discourse. 
Here, not only would the grammar have to be modified, but so would the semantic 
representation, to take into account indirect speech. In Discourse 1, this problem is not 
apparent: all verb tenses used are distinguished in our grammar because we limited 
ourselves to a subset of the data. 
Thus, the production of VP is more complex than what we have implemented and 
we have not completely identified all the rules for the selection of verb tense: indirect 
discourse is not implemented and we have not identified how modal information can 
be used to select forms such as the subjunctive and the conditional. But our approach 
is a good start and it could be extended by adding more systems and their selection 
rules, without changing the overall structure of the network. 
We can see from the approximate translations given in Table 9 that the systems 
for generating French and English verb tenses differ greatly. For English verb tenses, 
the method implemented in Nigel resorts to a recursive semantics involving temporal 
markers, as proposed by Halliday (Matthiessen and Bateman 1991). The purpose in 
this approach is to deal correctly with complex structures, such as will have been eating. 
Put simply, the idea is that each auxiliary reflects a relation between two temporal 
markers. This suggests a network that displays a recursive process. Thus, the phrase 
will have been going to eat would be represented semantically as to eat at a time that is in 
the future relative to another time that is in the past relative to a time that is in the future relative 
to speech time. 
This method may be adequate for English, since it seems to capture the recursive 
structure of verb tenses; but in French, this recursive structure is not found. Fur- 
thermore, nothing is said about how a deep generation process could produce the 
corresponding semantic structure with the intermediate temporal markers. In fact, we 
are not convinced that this could be easily done. Rather, we think that it is the overall 
structure that is selected for a particular usage. 
This completes our brief description of the generation of verb tenses. We have 
not completely solved the problem. In particular, we have chosen to put most of the 
problems pertaining to verb tense in the deep generation process, in order to facilitate 
their generation at the surface level. This approach greatly simplifies the process and 
our grammar could be easily completed to encompass all cases. Once the semantic 
demands are better understood, it should be easier to solve the problem of deep 
generation. 
7. Conclusion and Future Work 
In this paper, we have presented a method that has been successfully used to produce 
text conveying temporal information. Our method combines the principles of two 
theories: Kamp's Discourse Representation Theory, which guides the expression of 
temporal information, and Halliday's Systemic Functional Grammar, which provides 
a generation process controlled by a set of semantic choices, with the syntactic form 
resulting from these choices. We argued for the use of a conceptual structure, a Dis- 
course Representation Structure, combined with rhetorical principles and pragmatic 
information, and for its translation into a semantic structure that is easily realized syn- 
tactically. The deep generation process is hard to implement, mainly because of the 
difficulty in formalizing this information. Since we assume that the deep generation 
124 

References
1   James F. Allen, Maintaining knowledge about temporal intervals, Communications of the ACM, v.26 n.11, p.832-843, Nov. 1983 

2   Asher, Nicholas. (1993). Reference to Abstract Objects in Discourse. Kluwer Academics, Dordrecht. 

3   Bach, Emmon. (1986). The Algebra of Events. Linguistics and Philosophy, 9(1): 5--16. 

4   Berry, M. (1975). An Introduction to Systemic Linguistics, vol. 1 Structures and Systems. St. Martin Press, New York. 

5   Orna Berry , Mathias Hein , Ellon Littwitz, Introduction to Ethernet Switching, International Thomson Computer Press, Boston, MA, 1995 

6   Bras, M. (1990). Calcul des structures temporelles du discours. Thse de doctorat, Universit Paul-Sabatier. 

7   Bras, Myriam and Nicholas Asher. (1994). Le raisonnement non monotone dans la construction de la structure temporelle de textes en franais. In 9me congrs AFCET-RFIA. Paris. 

8   Bras, Myriam and Molins, Frdrique. (1993). "Adverbials of Temporal Location: Linguistic Description and Automatic Processing." In Sprache - Kommunikation - Informatik. Linguistiche Arbeiten 293, edited by J. Darski and Z. Vetulani. Max Niemeyer Verlag, Tubingen. 

9   Davidson, D. (1967). "The Logical Form of Action Sentences." In Essays on Action and Events, edited by D. Davidson. Clarendon Press. 

10   Dowty, D. (1979). Word Meaning and Montague Grammar. Reidel, Dordrecht. 

11   Dowty, D. (1982). Tenses, Time Adverbs, and Compositional Semantic Theory. Linguistics and Philosophy, 5(1): 23--55. 

12   Dowty, D. (1986). The Effects of Aspectual Class on the Temporal Structure of Discourse: Semantics or Pragramatics? Linguistics and Philosophy, 9: 37--61. 

13   Ehrich, Veronika. (1987). "The generation of tense." In Natural Language Generation: New Results in Artificial Intelligence, Psychology and Linguistics, edited by Gerard Kempen. Martinus Nijhoff Publishers, Boston, Dordrecht, 423--440. 

14   Fawcett, R. (1988). "Language Generation as Choice in Social Interaction." In Advances in Natural Language Generation---An Interdisciplinary Perspective, edited by Michael Zock and Gerard Sabah. Pinter Publishers, London, 27--48. 

15   Forster, D. (1989). Generating Temporal Expressions in Natural Language. In Proceedings, 11th Annual Conference of the Cognitive Science Society. 

16   Gagnon, M., and Lapalme, G. (1992). Un gnrateur de texte exprimant des concepts temporels. Technique et science informatiques, 11(2): 25--44. 

17   Gagnon, Michel and Bras, Myriam. (1994). "Discourse Interpretation and Time Representation." Technical report 94/54-r, IRIT. 

18   Halliday, M. (1985). An Introduction to Functional Grammar. Edward Arnold, London. 

19   Hovy, Eduard H. (1991). "Approaches to Planning of Coherent Text." In Natural Language Generation in Artificial Intelligence and Computational Linguistics, edited by W. Swartout and W. Mann. Kluwer Academics Publishers, Boston, 83--102. 

20   Kamp, H. (1979). "Events, Instants and Temporal Reference." In Semantics from different points of view, edited by R. Bauerle, U. Egli, and A. von Stechow. Springer Verlag, Berlin, 376--417. 

21   Kamp, H. (1981). Evnements, reprsentations discursives et rfrence temporelle. Langages, 64: 34--64. 

22   Lascarides, Alex, and Asher, Nicholas. (1993). Temporal Interpretation, Discourse Relations and Commonsense Entailment. Linguistics and Philosophy, 16: 437--493. 

23   Mann, W. (1983). "An Overview of the Nigel Generation Grammar." Technical report ISI-RR-83-113, USC/ISI. 

24   Mann, W. (1985). An Introduction to the Nigel Text Generation System. In Systemic Perspective on Discourse, vol. 1, edited by James D. Benson and William S. Greaves. Ablex, Norwood, 84--95. 

25   Mann, W., and Thompson, S. (1987). "Rhetorical Structure Theory: Description and Construction of Text Structure." In Natural Language Generation, edited by Gerard Kempen. Martinus Nijhoff, Dordrecht, 85--95. 

26   William C. Mann, Discourse structures for text generation, Proceedings of the 22nd annual meeting on Association for Computational Linguistics, p.367-375, July 02-06, 1984, Stanford, California 

27   Matthiessen, C. (1985). The Systemic Framework in Text Generation: Nigel. In Systemic Perspective on Discourse, vol. 1, edited by James D. Benson and William S. Greaves. Ablex, Norwood, 96--118. 

28   Matthiessen, C., and Bateman, J. (1991). Text Generation and Systemic-Functional Linguistics. Pinter, London. 

29   Maybury, Mark T. (1991). Topical, Temporal, and Spatial Constraints on Linguistic Realization. Computational intelligence, 7(4): 266--275. 

30   Kathleen R. McKeown, Discourse strategies for generating natural-language text, Artificial Intelligence, v.27 n.1, p.1-41, Sept. 1985 

31   Molins, F. (1990). Acceptabilit et interprtation des adverbiaux de localisation temporelle. Mmorie de D.E.A., Universit de Toulouse---Le Mirail. 

32   Reichenbach, H. (1947). Elements of Symbolic Logic. McMillan, New York. 

33   Verkuyl, H. J. (1989). Aspectual Classes and Aspectual Composition. Linguistics and Philosophy, 12: 39--94. 

34   Vlach, Frank. (1993). Temporal Adverbials, Tenses and the Perfect. Linguistics and Philosophy, 16: 231--283. 

35   Winograd, T. (1983). Language as a Cognitive Process. Addison Wesley. 
