A COMPOSITIONAL APPROACH TO THE TRANSLATION OF 
TEMPORAL EXPRESSIONS IN THE ROSRTTA SYSTEM 
Lisette hppelo 
Philips Research Laboratories 
Eindhoven, The Netherlands 
Abstract 
This paper discusses the translation of temporal 
expressions, in the framework of the machine tansla- 
tion system Rosetta. The translation method of 
Rosetta, the "isomorphic grammar method', is based on 
Montague's Compositionality Principle. It is shown 
that a compositional approach leads to a transparent 
account of the complex aspects of time in natural 
language and can be used for the translation of 
temporal expressions. 
O.Introductlon 
It is a well~known fact that the translation of 
temporal expressions in natural languages is not a 
simple mapping of verbal tenses. In (I) the Dutch 
Simple Present Tense is used while in (2) a Present 
Perfect Progressive Tense is the most appropriate 
tense to represent the time meanin~ of (i) in 
English. A more literal translation of the Dutch 
Simple Present is sometimes considered to be ill- 
formed, as Illustrated by (3). But in other cases 
that translation yields a perfect result, as shown by 
the pair (4)/(5). 
(i) Jan woont hier al 20 Jaar 
(Dutch: John-lives-here-already-2O-years) 
(2) John has been living here for 20 years 
(3) * John is living here for 20 years 
(4) Jan woont hier (John-lives-here) 
(5) John is living here 
In this paper I will deal with this kind of problem 
from the perspective of machine translation. More 
specifically, I will sketch a solution within the 
framework of the Rosetta machine translation project. 
In this project translation systems are developed 
between Dutch, English and Spanish, using the "iso- 
morphic grammar" method. According to this method, 
the gr~mnmars of the languages are attuned to each 
other in such a way that a sentence s is a transla- 
tion equivalent of a sentence s" if s and s" have 
slmJlar der~vational histories (el. Landsbergen 
(1984)). 
In section I I will give a short exposition of the 
isomorphic grammar method. Section 2 presents a 
theory of time in the Rosette framework. In section 3 
I will sketch isomorphic grammars for temporal ex- 
pressions and illustrate them by some examples. 
Possible extensions will be discussed in section 4. 
1. The Isonorphic Grammar Method 
1.2. H-Grammars 
M-grammars can be seen as a computationally viable 
and syntactically powerful variant of Montague Gram- 
mar. An M-grammar consists of three components: I) a 
syntactic component, 2) a morphological component and 
3) a semantic component. 
I) The syntactic component 
The syntactic component defines a set of S-trees 
(surface trees) whose leaves correspond to words, in 
surface order. An S-tree is an ordered tree with 
nodes which are labelled with syntactic categories 
and attribute-value pairs. The branches are labelled 
with syntactic relations, e.g. subject, object, etc. 
In the rest of this paper I will abbreviate them by 
specifying the top node and a characterization of the 
rest of the tree, as: 
CAT{attribute: value,...} (string) 
The syntactic component: defines S-trees by spec- 
ifying: 
(i) a set of basic S-trees 
(also called basic expressions) 
(ii) a set of syntactic rules. 
Starting from basic expressions larger expressions 
are formed by applying syntactic rules. 
A simplistic example is shown in the left part of 
figure I. The S-tree for the sentence the woman is 
singing is derived from the basic expressions woman 
and sing by applying rules R I and R . 
The process of making an expression can be repre- 
sented by a syntactic derivation tree (D-tree) with 
the (names of the) basic expressions at the terminal 
nodes and the names of the rules that are applicable 
at the nonterminal nodes. 
2) The morphological component 
The morphological component relates lexical S-trees 
(the leaves of the surface trees) to strings. So the 
syntactic and morphological components together de- 
fine sentences. 
I will ignore this component in this paper. 
3) The semantic component 
M-grammars obey the Compositionality Principle, which 
states that the meaning of an expression is a 
function of the meaning of the parts of that ex- 
pression. The basic S-trees and all other S-trees can 
be given a model theoretical interpretation: the 
basic expressions correspond to semantic values in a 
semantic domain and the rules to semantic operations. 
We can represent this in a semantic derivation tree 
that corresponds to the syntactic D-tree and that is 
labelled with the names of the meanings of the basic 
expressions at: the terminal nodes and the names of 
the semantic operations at the nonterminal nodes. 
(cf. figure i) 
M-grammars must satisfy certain conditions to allow 
for effective analysis next to generation. For more 
details the reader is referred to Landsbergen (1982, 
19~4). 
1.2. Translating with :Lsomorphle M-grammars 
The translation relation between two (or more) lan- 
guages is defined by attuning their grammars as 
follows: 
(i) For each basic expression of a grammar there is 
at least one corresponding basic expression of the 
other grammar with the same meaning. 
313 
(ii) For each syntactic rule of a grammar there is at 
least one corresponding syntactic rule of the other 
grammar, with the same meaning. 
The correspondence between rules is only required for 
"meaningful" rules. Syntactic transformations can he 
added for each language separately and do not occur 
in the derivation trees. 
In terms of derivation trees: for each syntactic 
derivation tree of a grammar there is at least one 
syntactic derivation tree of the other grammar with 
the same geometry and labelled with corresponding 
basic expressions and syntactic rules. These syn- 
tactic derivation trees correspond to the same seman- 
tic derivation tree and are called isomorphic deriva- 
tion trees. 
Two sentences are defined as translations of each 
other if they have the same semantic derivation tree 
and therefore corresponding derivation trees. (Note 
that the definition is given for isolated sentences 
on the basis of linguistic knowledge only: in fact a 
"possible translation" relation is defined, possible 
in some context.) 
Grammars that are attuned in this way are called 
isomorphic M-grammars if the corresponding sets of 
rules satisfy certain applicability conditions, such 
that for each well-formed syntactic derivation tree 
in a grammar there is at least one well-formed 
syntactic derivation tree in other grammars. (A 
derivation tree is well-formed if it defines a 
sentence, i.e. if the rules are applicable.) 
figure I 
English 17~ Dutch 
-----RI b2- ~ i B2 \[~ 
sing zi\[ngen 
~--bl BI "I I 
! ~Sn ~ICOUW 
synt. D-tree s6m. \])-tree synt. D-tree 
In figure 1 a simple unambiguous - example of 
isomorphic grammars for English and Dutch is given. 
The Rosetta translation systems are based on iso- 
morphic grammars. The semantic derivation trees func- 
tion as an intermediate language (IL). Generative and 
analytical components can be derived from the gram- 
mars: the analytical component maps a sentence of the 
source language into one or more semantic D-trees; 
the generative component maps a semantic D-tree into 
one or more sentences of the target language. 
In this paper the translation relation is discussed 
from a purely generative point of view. The semantics 
will be treated rather informally and I will es- 
pecially discuss the correspondence of the syntactic 
rules and show the development of parallel deriva- 
tions of sentences. 
2. A Time theory for Rosetta 
A linguistic analysis of temporal expressions, in 
particular an MT approach that presumes an inter- 
lingua, requires some level of representation that is 
314 
neutral with respect to the morpho/syntactic means 
for expressing time in natural language. As I prefer 
to adhere tc the current notions developed in the 
literature, I will discuss the relevant data in the 
perspective of the model-theoretical analyses put 
forward is e.g. Reichenbach (1947), Comrie (1976), 
Van Eynde e.ao (1985), Oversteegen and Verkuyl 
(1985), Bruce (1972), De Vuyst (1983). 
2.1. Temporal expressions 
Temporal expressions consist of: 
tense, a linguistic category which consists of 
morphological forms of the verb (e.g. worked, works) 
or of auxiliary verb forms in combination with 
certain morphological fomns of the verb which I call 
"periphrastic tenses" (e.g. has worked, is working; 
Spanish: e_st~ trabajando). 
time adverbials, liugnistie categories consisting 
of time adverbs, prepositional phrases or time con- 
junction phrases. 
2.2. The time model 
I will assume that all temporal expressions have a 
function in relating the event described by a sen- 
tence or a clause to a time model. Vor present 
purposes a simplified specification of the features 
of the mode\] I assume suffices. (For definitions see 
e.g. Bruce (1972)) 
This time model T is a partially ordered set of which 
the elements are called time points. The ordering 
relation is "<", meaning "earlier than". For this 
model we define a notion interval. Intervals are 
subsets of T without any gaps or branches. 
One of those points is called S, the moment of speech 
or narration. 
The "objects" to be located in T will be called 
events. An event is something that can be located in 
time, dependent on the temporal ingredients of the 
sentence that refers to the event. For example in 
order to establish the truth value of John has been 
living here for 20 years, we need to locate the event 
"John's living here" in T. 
An event is assumed to correspond to an interval E of 
T. We will say that the event is located in T When 
the relation between E and S can be established. This 
relation can be complex in the sense that more 
intervals than E and S can be involved. Such inter- 
vals are called reference intervals. 
Intervals can be characterized by properties indica- 
ting e.g. the "length" or "duration" of an interval, 
a particular relation to S or a part of the calendar. 
These properties are expressed by adverbials or 
special (auxiliary) verbs. 
The event which corresponds to the interval E has 
temporal properties which are often called 
"Aktionsart" in the literature. Both the main verb 
and its arguments of the clause which expresses the 
event may play a role in the determination of the 
Aktionsart (cf. Verkuyl (1972)). 
Usually four types of Aktionsart are be distinguished 
(ef. e.g. De Vuyst (1983)) as is illustrated in the 
following examples: 
(6) This book belongs to me (stative) 
(7) John is working (activity) 
(8) John wrote a letter (accomplishment) 
(9) He reached the end of the street (achievement) 
The event in (6) is durative; it can take place at an 
arbitrarily long interval. In (7) the event is 
durative, but it cannot be claimed to take place only 
at a time point (i.e. a minimal interval) in the 
model; it :Is dynamic which implies some progress or 
change. The event in (8) is terminative, because the 
result or end is indicated; it can be looked at from 
the "outside" as a unit, but it cannot be claimed to 
take place at a time point in the model; it is 
dynamic when looked at from the "inside" (activity). 
In (9) the event is terminative, because the end or 
result is indicated but Jt is also claimed to take 
place at a time point in the model and therefore also 
called momentary° 
There are two important types of relations between 
intervals: 
I) the deictic relation: relation between a reference 
interval and S 
2) the aspectual relation: relation between E and a 
reference interval 
These relations are expressed by morphological and 
periphrastic tense. 
2.3. Time in Rosetta 
For definin~ the translation of temporal expressions 
in the Rosette framework, we have to write isomorphic 
compositional grammars for them, which boils down to: 
a) specifying for each language: 
(i) temporal expressions (time adverbials etc.) ex- 
pressing properties of time intervals, 
(ii) syntactic rules (e.g. tense rules) that indicate 
how temporal expressions must and can be combined, 
expressing relations between those intervals, 
b) attuning these expressions and rules of the 
languages involved to each other ~n the way described 
in section \]. 
Before specifying these grammars in section 3 I will 
briefly discuss the motives for the strategy followed 
in these grammars and for the particular choice of 
reference intervals. 
I. Tenses and adverbials cannot be translated inde- 
pendently: 
- Sometimes the translation of a tense is only 
correct if it occurs with a certain time adverbial. 
Consider for example the pair (i0)/(ii): 
(i0) Jan werkt bier al 3 jaar (Pres. Tense) 
(11) John has been working here for 3 years 
(Pres. Perf. Prog. Tense) 
The corresponding tense of the Dutch Present Tense in 
the context of the time adverbial al 3 jaar in (i0), 
is in English a Present Perfect Progressive Tense as 
in (Ii). But without that time adverbial the trans- 
lation is different as is sho~1 in the examples 
(12) - (14): 
(12) Jan werkt (Pres. Tense) 
(13) John is worklng (Pres. Prog. Tense) 
(14) *John has been working (Pres. Perf. Prog. Tense) 
- Adverbla\]s are not always translated into adverb- 
ials. Consider for example the pair (15)/(16): 
(15) English: He has just arrived. 
(16) Spanish: E1 acaba de llegar. 
In (15) the adverb Just expresses the "near past", 
but in (16) a special verb acabar de which could be 
considered as a part of some perifrastic tense is 
used. 
Those examples are an indication that the grammars 
for temporal expresions should take Into accoun\[ 
tenses and time adverbials together. 
2. Van Eynde et al. (1985) give a specification of 
time meaning representation for machine translation, 
based on a time model with three intervals, E, R and 
S, which results in time meaning representations that 
do take into account time adverbials. 
In their time model, however, the set of aspectual 
relations, relations between E and R, contains next 
to tile retrospective relation (E before R) the 
imperfective relation (E contains R). This seems 
inadequate in view of sentences that can have both an 
imperfective and a retrospective aspectual relation 
as in e.g. (10) and (ll). 
The complexit:y of the aspectnal relation is also 
recognized in Maegaard (1982) where a special value 
CONTINUOUS for the attribute RETROSPECTVE was intro- 
duced for the translation of verbal tenses. 
Krauwer and Des Tombe (1985) make similar observa- 
tions. 
I propose therefore that in the Rosetta framework 
I) a time meaning representation obligatorily will 
contain an aspeetual relation, i.e. a relation be- 
tween E and a reference interval R~, which will be 
called perfective if E is a subset ~ R E and 
imperfee--t~ve~f~R~ is a subset of E , and 
~) t~at .it can t"bptionally contain a retrospective 
relation between R E and a time p~int RS, which is 
sone "local l×}Int of evaluation" , meaning that ~( 
lasts until R S. 
This R_ is an arbitrary point of some reference 
Interval. Consider for example- 
(17) Yesterday John had been living there for 3 years 
The 3 years (RE) last until some point during 
yesterday (Rs). 
If there is no retrospective relation, R S will he an 
arbitrary point of R F. 
Between R S and S th~ d~-~ctic relation is specified: 
Past (R e 'is before S), Present ( R S is simultaneous 
W~ s)%r ~:,~F~r~ ~H s is ~t~T s). 
So a time meaning representation of an event in 
Rosetta will eonsist of: 
- properties of E and R E 
an aspectual relation"between E and R E 
optionally a retrospective relation between R E and 
R S 
- a' deictic relation between R S and S 
In the next section grammars for temporal expressions 
will be discussed which start with a clause and apply 
rules that will first specify the properties for E, 
then the aspectual relation and the properties for 
R E, then optionally the retrospective relation and 
finally the deictic relation. 
3. Isounorphle. grammars for temporal expressions 
3.1. Corresp.nding rules for temporal expressions 
To achieve i,~omorphy of grammars for temporal ex- 
pressions, corresponding rules for the languages of 
the system have to be written as was explained in 
section I. These rules are applied to a clause which 
consists of a verb, its arguments and an attribute 
Aktionsart, the value of which has been specified 
during the composition of the clause. The result of 
the application of the rules is a clause with 
specified tense forms, auxiliaries and adverbials. 
Semantically, properties of and relations between 
315 
time intervals and the event are specified. The rules 
have one argument~ a clause, or two arguments, a 
clause and an adverbial or an auxiliary verb that we 
wish to introduce categorematically. 
I will distinguish five classes of rules. The rules 
are either obligatory (OB)~ meaning that exactly one 
of this class of rules is applied, or optional (OP). 
The rules will be applied in the following order 
(from a generative, compositional point of view). 
I. Aktionsart rules (OP): the application of these 
rules results in a new clause with a different 
Aktionsart value caused by some (auxiliary) verb or 
adverbial that is inserted into the clause. 
If. duration rules (OP): These rules insert a dur- 
ation adverbial into the clause. 
III. aspect rules (OB): These rules insert a refer- 
ence adverbial and specify the aspectual tense forms 
(perfective or imperfectlve) of the verbs. 
IV. retrospective rules (OP): These rules are applied 
to a clause that contains some retrospective refer- 
ence adverbial. It inserts another, non-retrospectlve 
reference adverbial and adds, if necessary, auxiliary 
verbs and/or adverbials. 
V. deictic rules (OB): These rules determine the 
deictic tense form of the verbs in the clause. 
In 3.3. the rules will be discussed in more detail. 
3.2. The S-trees 
A clause is represented as an S-tree with a top node 
CL that has the following temporal attributes and 
corresponding value sets in all languages: 
Aktionsart: {stative, activity, 
accomplishment, achievement} 
aspect: {imperfective, perfective, unmarked} 
deixis: {present, past, future, unmarked} 
retrospeetivity: {-retro, q-retro~ 
(the underlfned value is tile initial value). 
A clause contains a VERB node with attribute-value 
pairs concerning the verb form, which may differ over 
languages. 
A clause represents an event with time interval E. 
"lq~e other temporal expressions may be of various 
categories: e.g. ADVP, PP, NP or CONJP (a time 
conjunction and a clause). 
They are marked at the topnode for temporal prop- 
erties by the following attributes: 
class: {duration, reference} 
deixis: {present, past, future~ unmarked} 
aspect: {perfective, imperfective} 
retrospectivity: {+retro, -retro} 
- Adverbials of the duration class will always have 
the value unmarked for their deixis and -retro for 
their retrospectivity attribute. They indicate a 
property of the interval E. Perfective duration 
adverbials specify the duration of the event, imper- 
fective the duration of an interval during which the 
event takes place. For example: 
PP{class: duration, aspect: imperfective, 
deixis: unmarked, retrospectivity: -retro} 
(in three hours) 
- Reference adverbials indicate properties of inter- 
vals R. If their deixis attribute has the value 
unmarked, they are called absolute, indicating that 
there are no restrictions on the relations of the 
interval with S, otherwise they are called deictic, 
indicating that the interval has a certain relation 
with S. If their retrospective attribute has the 
value +retro, they are called retrospective, indi- 
cating that the interval has the relation until to 
the reference point R S. For example: 
316 
ADVP{class: reference, aspect: imperfective, 
deixis: past, retrospeetivity: -retro} 
(yesterday) 
Absence of adverbials 
Clauses do not always contain explicit adverbials: 
a) In case of isolated clauses without reference 
adverbials we will assume an abstract deictic refer- 
ence adverbial which indicates the moment of speech 
S: 
REF{class: reference; deixis: present; 
aspect: perfective; retrospectivity: -retro}() 
In general, clauses or sentences occur in texts and 
reference adverbials can have scope over subsequent 
clauses. For those latter clauses we will assume that 
they have abstract anaphoric reference adverbials 
with the properties of their antecedents. 
b) In case a reference interval is indicated that has 
the property that it ranges over the whole time axis 
until some reference point RS, we assume an abstract 
adverbial: 
PAST {class: reference; deixis: unmarked; 
aspect: imperfective; 
retrospectivity: +retro} () 
3.3. The rules 
I will now give an informal description of each type 
of rule containing an account of syntactic and 
semantic aspects and the differences between Dutch, 
English and Spanish. 
I. Aktionsart rules (OP). ~ley change the Aktionsart 
and insert (complex) auxilisry verbs or adverbs, 
and/or determine the form of the verb. Semantically, 
a new event is derived from the original event. 
For exmnple, an accomplishment event can be trans- 
formed into a non-terminatlve event. Compare: 
Eng: (18) CL (John write a letter) --> 
CL (,John be writing a letter) 
Du: (19) CL (Jan een brief schrijven) --> 
CL (Jan een brief aan het schrijven zijn) 
This is sometimes called the "locative tense"; the 
rules insert: 
in Dutch: aan het VERB{fonn: Infinitief}zijn 
in Spanish: estar VERB{form: gerundio} 
in English: be VERB f\[~rm: ingform} 
They change the Aktionsart value to stative (due to 
the auxiliary). 
II. duration rules (OP). These rules are applied to 
an S-tree with a certain Aktionsart value and a 
duration adverbial with a certain aspect value that 
is inserted in the clause. The aspect value of the 
clause will now be perfective. 
This rule applies the property denoted by the adverb- 
ial, to the interval E0 
For example: 
(20) CL (John write for three hours) 
III. aspect rules (OB). Rules with two arguments: I) 
a clause with a certain Aktionsart and aspect and 2) 
a (possibly anaphoric) reference adverbial with a 
certain aspect. They determine the imperfeetive and 
perfective verb forms, sometimes with auxiliaries. 
The aspectual verb forms can differ over languages. 
English seems to have perfective simple tense forms; 
the imperfective forms are composed with the auxili- 
ary be. Spanish has clear imperfective and perfective 
past tense forms. In Dutch the verb fo~1 does not 
seem to distinguish between imperfective and perfec- 
tive. 
The aspect value of the clause unmarked has to be 
changed into perfective or imPerfective. The refer- 
ence adverbial is inserted into the S-tree. 
Semantically, the property denoted by the adverbial 
is applied to the interval R~ and the relation 
between the intervals E and R E i~ expressed. 
Example: 
(21) CL (Ayer Juan ley6 un libro.) (perfective) 
(Yesterday-John-read-a-book) 
(22) CL (Ayer Juan lela un llbro.) (imperfeetive) 
(Yesterday-John-was-reading-a-book) 
IV. retrospeetivity rules (OP). Rules that have two 
arguments: 1) a clause with a certain aspect and a 
reference adverbial that has the retrospectivlty 
value +retrn (this may be PAST), and 2) a (possibly 
anaphoric) reference adverbial that has the retro- 
spectivity w11ue -retro. They insert auxiliary verbs 
such as have (hebb~jn (Dutch), haber (Spanish)), 
if necessary. In English have must always be in- 
serted. In Dutch and Spanisll it is obligatory if the 
aspect value is perfective. So these languages have 
two rules: one for perfective and one for imperfee- 
tive clauses. The reference adverbial is inserted in 
the S-tree. 
The rules that insert have just (English) zojuist 
hebben/zijn (Dutch) and acabar de (Spanish), under 
the condition that the clause contains the abstract 
adverbial PAST and the aspect value perfective, to 
express near-retrospectivity, be\]ong to this class. 
Semantically, these rules apply the property denoted 
by the adverbial to R S and express the relation until 
or "near~until" of RE,' (indicated by the retrospec~ 
adverbial) to R . S 
Ex am pl e : 
(23) CL (REF John have just read this book.) 
(24) CL (REF Juan acabar de leer este libro.) 
(25) CL (REF Jan dit boek zojnist gelezeu hebben) 
V. deietic rules (OB). These rules are applied to a 
clause with a reference adverbial. (that can be 
anaphorie), \[nserted by rules of type IIl or IV. They 
determine present, past and future forms of the 
verbs, sometimes with insertion of an auxiliary verb 
form, for example will for future in English, after 
checking if the deictic ~alue of the reference 
adverbial is compatible. 
Semantically, the relation between Ro (a subset of 
the interval indicated by the reference adverbial in 
the clause) and S is expressed. The deietic relations 
are: present (R~ is simultaneous with S), past (R S is 
befor~d f~ture (R~ is after S). In these rHles 
the abstract adverbi-~als'~will be deleted. 
Example: 
(26) John read a book yesterday 
(27) *John has read a book yesterday 
(28) John had read a book yesterday 
3.4. Examples 
I will now give some examples of parallel derivations 
of sentences with temporal expressions that are 
translation equivalents. I will leave out irrelevant 
specification of nodes. 
Example: (29) John has been writing for 2 hours 
Example application of syntactic rules for English 
RI: (imperfective aspect rule) 
CL{Aktionsart: activity, aspect:unmarked, 
deixis: unmarked, retrospectivity: -retro} 
(John write) 
+ PP{deixis: unmarked, aspect: perfective, 
retrospectivity: +retro~ class: reference} 
for 2 hours) 
--> CL ..., aspect: perfective,...} 
(John be writing for 2 hours) 
R2: (retrospective rule) 
CL{..., retrospectivity: -retro} 
(John be writing for 2 hours) 
+ REF{deixis: present, aspect: perfective, 
retrospectivity: -retro, 
class: reference}() 
--> CL{..., retrospectivity: +retro} 
(John have been writing for 2 hours REF) 
R3: (present deictic rule for finite clause) 
~{deixis: unmarked,...} 
(John have been writing for 2 hours) 
--> CL{deixis: present,...} 
(John has been writing for 2 hours) 
figt~e 2 
synt. D-tree s~t. D-tree 
PP (\], ADVP CL 
John write Jan schrijven 
Figure 2 contains the isomorphic syntactic D-trees 
for the English sentence (29) and its Dutch transla- 
tion equivalent : 
(30) Jan schr~jft al 2 uur 
(Dutch: John=writ es-already-2-hour s) 
The Dutch rules differ from the English ones in that 
R "1 does not insert an auxiliary for the imperfective 
fo~m as in English. It results in: Jan al 2 uur 
schrijven. R', differs from R also with respect to 
~e introoduct~on of the auxiliary verb. lu Dutch it 
is not necessary to insert hebben/zijn if the clause 
has imperfective aspect. R" 2 results in: Jan REF al 2 
uur sebrijven. R" 3 is like R3, but in fi~te~clauses 
the las-t~ ve--rh is placed in second position in Dutch: 
Jan schrijft al 2 uur. 
flgure 3 
i~3 ~'3 
R" 2 
pp._  l\ / \R 
John write Jan sehrijven 
(30) is only one of the possible translations of 
(29): (29) is ambiguous. The be writing can also be 
due to some "Aktionsart rule"--~4 ) corresponding to 
aan het schrijven zijn (R'4). The other Dntch trans- 
lation is: 
317 
(31) Jan is al 2 uur aan het schriJven 
(John-is-already-2-hour s-on~the-wr it ing) 
The isomorphic D-trees for (29) and (31) are shown in 
figure 3. 
But for 2 hours is m, biguous too and can also be a 
perfective duration adverbial corresponding to the 
Dutch 2 uur. The duration rules (R~ and R'5) insert 
them. ~he ~lause will now have perf~ectlve aspect and 
the retrospective adverbial PAST will be introduced 
by the perfective aspect rule (R~ and R'6). In Dutch 
the retrospective rule R" 7 (different f?om R" 2 but 
also corresponding to the English R2) is now applied: 
the one for a clause with perfective aspect that 
inserts the auxiliary hebben/zijn. The resulting 
Dutch translation equivalent of (29) is: 
(32) Jan is 2 uur aan bet schrijven geweest 
(Dutch : John~has-2-hour s-on-the-wr it ing-been) 
The isomorphic derivation trees are shown in figure 
4. 
fig~e 4 
R 2 " 
'~R 4 NP~R" 4 
John write 
The conditions in the rules 
unacceptable combinations, e.g.: 
(33) *John died for a while 
filter out certain 
(34) *John was working in three hours 
(33) is ruled out because for a while is a perfective 
duration adverbial that cannot combine with achieve- 
ments or accomplishments and (34) because in three 
hours is an imperfective duration adverbial that 
cannot combine with activities and statives. 
4. Concluding remarks 
It is not possible to treat all temporal expressions 
and all translation problems with respect to time in 
this paper, but I have sketched a solution as to how 
to treat them in the Rosetta framework. I expect that 
other "aspectual forms" such as inchoative, termina- 
tive etc. can be added at the "Aktionsart level" or 
"aspect level" and that the current approach, which 
allows for translation of adverbs into auxiliary 
verbs or combinations of them (and vice versa) will 
be sufficient to cope with them. The time theory 
presented here should of course he embedded in a 
discourse theory about time. The anaphoric properties 
should be seen as a start. Moreover, the informal 
specification of the time model should be given a 
more formal account. Other topics that should be 
elaborated are quantificational, scope, frequency and 
habitual aspects, interaction of modality in the 
future time expressions and the time relations be- 
tween matrix clauses and (intensional) complement 
clauses or relative clauses. 
Acknowledgements 
The author would like to thank all members of the 
Rosetta team, particularly Jan Landsbergen and 
Franciska de Jong, for their helpful comments on 
earlier versions of this paper. 
This work was supported in part by a grant from the 
Nederlandse HerstructureringsmaatschappiJ (NRHEM). 
Notes 
I) I use the definition of "aspect" of Comrie (1976), 
but I will take his term "situation" as referring to 
the interval E in relation to the reference interval 
R E and consider only two possibilities: "perfective" 
and "imperfective" aspect. The notion "perfect 
aspect" which I call 'retrospective aspect" should be 
accounted for by other relations. Probably~ by rela- 
ting R E to R S with the relation "until". 
2) This point resembles the point S" in Oversteegen and Verkuyl (1985). 
3) It is possible that some languages have complex 
deictic rules that indicate for example a "past event 
with present relevance". This is similar to a "pre- 
sent with retrospective PAST", but differs In that 
the interval R E is not until, but before R e. At the 
moment I will Ignore the e~-xtra meaning" ~nd treat 
them with a deictic past rule, because this phenom- 
enon seems to \]lave to do with more pragmatic factors 
such as "distance with respect to the event from the 
speaker's viewpoint" or knowledge of the world like 
the present existence of persons or objects. 

References 

Bruce, B., (1972): "A model for temporal references 
and its application in a question answering program'. 
In: Artificial Intelligence 3, \] -25. 

Comrie, B., (1976): Aspect: an Introduction to the 
Study of Verbal Aspect and Rela--~d Problems.~am- 
nlvers ty ress,-~r~-~g~. 

Dowty, D.R., (1979): Word Meaning and Montague 
Grammar. Reidel Vol. 7q--. ~~. 

Krauwer, S. and L. des Tombe, (1985): "Composi- 
tionality'. In: Dehaan, G. and W. Zonneveld (eds.): 
Formal Parameters of Generative Grammar. I. Yearbook 

Landsbergen, J., (1982): "Machine Translation based 
on Logically Isomorphic Montague Grammars'. In: J. 
Horecky (ed.), COLING 82 North Holland, p. 175-182. 

Landsbergen, J., (1984): "Isomorphic Grammars and 
their use in the Rosetta Translation system'. Paper 
presented at the Tutorial on Machine Translation, 
Lugano. Philips Research M.S. 12.950, to appear in: 
King, M. (ed.), Machine Translation: The State of the 
Art, Edinburgh Un-ivTr-s~\[Ty Press. 

Maegaard, B., (1982): "The Transfer of Finite Forms 
in a Machine Translation System'. In: Abstracts 
COLING 82, p. 190-194. 

Oversteegen, L. and H.J. Verkuyl, (1985): De 
temporele zinsstructuur van bet Nederlands: twee 
tijdsbanden'. In GLOT. 

Reichenbach, H., (1947): Elements of Symbolic Logic. 
University of California Press, 

Van Eynde, F., L. des Tombe and F. Maes (1985): 
"Specification of time meaning for machine translation'. Proceedings ACL conference, Geneva. 

Verkuyl, H.J., (1972): On the Compositional Nature of 
the Aspects. Dordrecht. 

Vuyst, J. de, (1983): "A Semantic Analysis of Tem- 
poral Elements in Dutch and English'. Diss. RUG, 
Groningen. 
