Telicity as a Cue to Temporal and Discourse Structure in 
Chinese-English Machine Translation* 
Mari Olsen David Traum 
Microsoft U Maryland 
molsen@microsoft.com traum@cs.umd.edu 
Carol Van Ess-Dykema 
U.S. Department of Defense 
carol@umiacs.umd.edu 
Amy Weinberg 
U Maryland 
weinberg@umiacs.umd.edu 
Ron Dolan 
Library of Congress 
rdolan@cfar.umd.edu 
Abstract 
Machine translation between any two s re- 
quires the generation of information that is implicit 
in the source . In translating from Chinese 
to English, tense and other temporal information 
must be inferred from other grammatical and lex- 
ical cues. Moreover, Chinese multiple-clause sen- 
tences may contain inter-clausal relations (temporal 
or otherwise) that must be explicit in English (e.g., 
by means of a discourse marker). Perfective and im- 
perfective grammatical aspect markers can provide 
cues to temporal structure, but such information is 
not present in every sentence. We report on a project 
to use the \]exical aspect features of (a)te\]icity re- 
flected in the Lexical Conceptual Structure of the 
input text to suggest tense and discourse structure 
in the English translation of a Chinese newspaper 
corpus. 
1 Introduction 
It is commonly held that an appropriate interlingua 
must allow for the expression of argument relations 
in many s. This paper advances the state of 
the art of designing an interlingua by showing how 
aspectual distinctions (telic versus atelic) can be de- 
rived from verb classifications primarily influenced 
by considerations of argument structure, and how 
these aspectual distinctions can be used to fill lexical 
gaps in the source  that cannot be left un- 
specified in the target . Machine translation 
between any two s often requires the gen- 
eration of information that is implicit in the source 
. In translating from Chinese to English, 
tense and other temporal information must be in- 
ferred from other grammatical and lexical cues. For 
example, Chinese verbs do not necessarily specify 
whether the event described is prior or cotempora- 
neous with the moment of speaking. While gram- 
matical aspect information can be loosely associated 
with time, with imperfective aspect (Chinese ~ zai- 
and ~ .zhe) representing present time and perfec- 
tiv e (Chinese Tle) representing past time, (Chu, 
* We gratefully acknowledge DOD support for this work 
through contract MDA904-96-R-0738 
1998; Li and Thompson, 1981), verbs in the past 
do not need to have any aspect marking distinguish- 
ing them from present tense verbs. This is unlike 
English, which much more rigidly distinguishes past 
from present tense through use of suffixes. Thus, in 
order to generate an appropriate English sentence 
from its Chinese counterpart, we need to fill in a 
potentially unexpressed tense. 
Moreover, Chinese multiple-clause sentences may 
contain implicit relations between clauses (temporal 
or otherwise) that must be made explicit in English. 
These multiple-clause sentences are often most nat- 
urally translated into English including an overt ex- 
pression of their relation, e.g., the "and" linking the 
two clauses in (1), or as multiple sentences, as in (2)). 
(1) 1 9 65 ~ ~ , ~ ,~ 
1965 year before , our_country altogether 
only have 30 ten_thousand ton de 
~t ~2 , ~ ~ ~8 
shipbuilding capacity , year output is 8 
ten_thousand ton 
Before 1965 China had a total of only 300,000 
tons of shipbuilding capacity and the annual 
output was 80,000 ~ons. 
(2)~ 8~ ~ ~ ~d Y 
this 8 ten_thousand ton actually include asp 
517 cl , ship de tonnage is very low de 
This 80,000 tons actually included 517 ships. 
Ship tonnage was very low. 
In our NLP applications, we use a level of linguis- 
tic structure driven by the argument-taking proper- 
ties of predicates and composed monotonically up to 
the sentence level. The resulting Lexical Conceptual 
Structures (LCS) (3ackendoff, 1983), is a - 
neutral representation of the situation (event or 
state), suitable for use as an interlingua, e.g., for 
machine translation. The LCS represents predicate 
argument structure abstracted away from - 
specific properties of semantics and syntax. The 
34 
primitives of the interlingua provide for monotonic 
composition that captures both conceptual and syn- 
tactic generalities (Dorr et al., 1993) among lan- 
guages. 1 The strength of the representation derives 
from the cross-linguistic regularities in the lexical se- 
mantics encoded in the LCS. The syntactic hierarchy 
(subject, object, oblique) is mirrored in the LCS hi- 
erarchy: for example THEMES are arguments of the 
LCS predicate, and AGENTS are arguments of the 
theme-predicate composition. Syntactic divergences 
(whether the object precedes or follows the verb, for 
example) are represented in  specific lin- 
earization rules; lexical divergences (whether the lo- 
cation argument is encoded directly in the verb, e.g. 
the English verb pocket or must be saturated by an 
exterfial argument) are stated in terms of the pieces 
of LCS struct-ure in the lexicon. SententiM repre- 
sentations derive from saturating the arguments re- 
quired by the predicates in the sentence. 
LCS rePresentations also include temporal infor- 
mation, where available in the source : re- 
cent revisions include, for example (Dorr and Olsen, 
1997a) standardizing LCS representations for the as- 
pectual (un)boundedness ((A)TELICITY) of events, 
either lexically or sententially represented. Although 
at present the LCS encodes no supra-sentential dis- 
course relations, we show how the lexical aspect in- 
formation may be used to generate discourse co- 
herence in temporal structure. Relations between 
clauses as constrained by temporal reference has 
been examined in an LCS framework by Dorr and 
Gaasterland (Dorr and Gaasterland, 1995). They 
explore how temporal connectives are constrained 
in interpretation, based on the tense of the clauses 
they connect. While overt temporal connectives are 
helpful when they appear, our corpus contains many 
sentences with neither tense markers nor tense con- 
nectives. We must therefore look to a new source of 
information. We rely on the lexical information of 
the verbs within a sentence to generate both tense 
and temporal connectives. 
Straightforward LCS analysis of many of the 
multi-clause sentences in our corpus leads to vio- 
lations of the wellformedness conditions, which pre- 
vent structures with events or states directly modi- 
fying other events or states. LCS, as previously con- 
ceived, prohibits an event or state from standing in a 
modifier relationship to another event or state, with- 
out mediation of a path or position (i.e., as lexically 
realized by a preposition). This restriction reflects 
the insight that (at least in English) when events and 
states modify each other, the modification is either 
implicit, with the relevant events and states in sepa- 
rate sentences (and hence separate LCSs), as in the 
1 LCS representations in our system have been created for 
Korean, Spanish and Arabic, as well as for English and Chi- 
nese. 
first sentence below, or explicit in a single sentence, 
as in the second sentence below. Implicit event-state 
modification (sentence 3) is prohibited. 
* Wade bought a car. He needed a way to get to 
work. 
* Wade bought a car because he needed a way to 
get to work. 
* * Wade bought a car he needed a way to get to 
work. 
It is exactly these third type that are permitted 
in standard Chinese and robustly attested in our 
data. If the LCS is to be truly an interlingua, we 
must extend the representation to allow these kinds 
of sentences to be processed. One possibility is to 
posit an implicit position connecting the situations 
described by the multiple clauses. In the source lan- 
guage analysis phase, this would amount to positing 
a disjunction of all possible position relations im- 
plicitly realizable in this . Another option 
is to relax the wellformedness constraints to allow 
an event to directly modify another event. This not 
only fails to recognize the regularities we see in En- 
glish (and other ) LCS structures, for Chi- 
nese it merely pushes the problem back one step, 
as the set of implicitly realizable relations may vary 
from  to  and may result in some 
ungrammatical or misleading translations. The sec- 
ond option can be augmented, however, by factoring 
out of the interlingua (and into the generation code) 
-specific principles for generating connec- 
tives using information in the LCS-structure, proper. 
For the present, this is the approach we take, us- 
ing lexical aspectual information, as read from the 
LCS structure, to generate appropriate temporal re- 
lations. 
Therefore not only tense, but inter-sentential dis- 
course relations must be considered when generating 
English from Chinese, even at the sentence level. We 
report on a project to generate both temporal and 
discourse relations using the LCS representation. In 
particular, we focus on the encoding of the lexical as- 
pect feature TELICITY and its complement ATELIG- 
ITY to generate past and present tense, and corre- 
sponding temporal relations for modifying clauses 
within sentences. While we cannot at present di- 
rectly capture discourse relations, we can garner as- 
pectual class from LCS verb classification, which in 
turn can be used to predict the appropriate tense for 
translations of Chinese verbs into English. 
2 Use of Aspect to Provide 
Temporal Information 
We begin with a discussion of aspectual features of 
sentences, and how this information can be used to 
provide information about the time of the situations 
35 
presented in a sentence. Such information can be 
used to help provide clues as to both tense and rela- 
tionships (and cue words) between connected situa- 
tions. Aspectual features can be divided into gram- 
matical aspect, which is indicated by lexical or mor- 
phological markers in a sentence, and lexical aspect, 
which is inherent in the meanings of words. 
2.1 Grammatical aspect 
Grammatical aspect provides a viewpoint on situa- 
tion (event or state) structure (Smith, 1997). Since 
imperfective aspect, such as the English PROGRES- 
SIVE construction be VERB-ing, views a situation 
from within, it is often associated with present 
or contemporaneous time reference. On the other 
hand, perfective aspect, such as the English have 
VERB-ed, Views a situation as a whole; it is there- 
fore often associated with past time reference ((Com- 
rie, 1976; Olsen, 1997; Smith, 1997) cf. (Chu, 
1998)). The temporal relations are tendencies, 
rather than an absolute correlation: although the 
perfective is found more frequently in past tenses 
(Comrie, 1976), both imperfective and perfective co- 
occur in some  with past, present, and fu- 
ture tense. 
In some cases, an English verb will specify tense 
and/or aspect for a complement. For example, con- 
tinue requires either an infinitive (3)a or progressive 
complement (3)b (and subject drop), while other 
verbs like say do not place such restrictions (3)c,d. 
(3) a. Wolfe continued to publicize the baseless 
criticism on various occasions 
b. Wolfe continued publicizing the baseless 
criticism on various occasions 
c. Wolfe continued publicizing the baseless 
criticism on various occasions 
d. He said the asia-pacific region already be- 
came a focal point region 
e. He said the asia-pacific region already is be- 
coming a focal point region 
2.2 Lexical aspect 
While grammatical aspect and overt temporal cues 
are clearly helpful in translation, there are many 
cases in our corpus in which such cues are not 
present. These are the hard cases, where we must 
infer tense or grammatical aspectual marking in the 
target  from a source that looks like it pro- 
vides no overt cues. We will show however, that 
Chinese does provide implicit cues through its lex- 
ical aspect classes. First, we review what lexical 
aspect is. 
Lexical aspect refers to the type of situation de- 
noted by the verb, alone or combined with other 
sentential constituents. Verbs are assigned to lexical 
aspect classes based on their behavior in a variety of 
syntactic and semantic frames that focus on three as- 
pectual features: telicity, dynamicity and durativity. 
We focus on telicity, also known as BOUNDEDNESS. 
Verbs that are telic have an inherent end: winning, 
for example, ends with the finish line. Verbs that are 
atelic do not name their end: running could end with 
a distance run a mile or an endpoint run to the store, 
for example. Olsen (Olsen, 1997) proposed that as- 
pectual interpretation be derived through monotonic 
composition of marked privative features \[÷/0 dy- 
namic\], \[.4-/0 durative\] and \[-t-/0 telic\], as shown in 
Table 1 (Olsen, 1997, pp. 32-33). 
With privative features, other sentential con- 
stituents can add to features provided by the verb 
but not remove them. On this analysis, the \[.-I-du- 
rative, +dynamic\] features of run propagate to the 
sentence level in run ~o the store; the \[÷telic\] feature 
is added by the NP or PP, yielding an accomplish- 
ment interpretation. The feature specification of this 
¢ompositionally derived accomplishment is therefore 
identical to that of a sentence containing a telic ac- 
complishment verb, such as destroy. 
According to many researchers, knowledge of lex- 
ical aspect--how verbs denote situations as devel- 
oping or holding in time-=may be used to interpret 
event sequences in discourse (Dowty, 1986; Moens 
and Steedman, 1988; Passoneau, 1988). In particu- 
lar, Dowty suggests that, absent other cues, a relic 
event is interpreted as completed before the next 
event or state, as with ran into lhe room in 4a; in 
contrast, atelic situations, such as run, was hungry 
in 4b and 4% are interpreted as contemporaneous 
with the following situations: fell and made a pizza, 
respectively. 
(4) a. Mary ran into the room. She turned on her 
walkman. 
b. Mary ran. She turned on her walkman. 
c. Mary was hungry. She made a pizza. 
Smith similarly suggests that in English all past 
events are interpreted as telic (Smith, 1997) (but cf. 
(Olsen, 1997)). 
Also, these tendencies are heuristic, and not abso- 
lute, as shown by the examples in (5). While we get 
the expected prediction that the jumping occurs af- 
ter the explosion in (5)(a), we get the reverse predic- 
tion in (5)(b). Other factors such as consequences of 
described situations, discourse context, and stereo- 
typical causal relationships also play a role. 
(5) a. The building exploded. Mary jumped. 
b. The building exploded. Chunks of concrete 
flew everywhere. 
36 
Aspectual Class 
State 
Activity 
-~Accomplishment 
Achievement 
Telie Dynamic Durative 
+ 
+ ,, , + __ run , paint 
+ + + 
+ + 
Examples 
know, have 
destroy 
notice, win 
Table 1: Lexical Aspect Features 
3 Aspect in Lexical Conceptual 
Structure 
Our implementation of Lexical ConceptuM Struc- 
ture (Dowty, 1979; Guerssel et al., 1985)--an 
augmented form of (Jackendoff, 1983; Jackendoff, 
1990)--permits lexical aspect information to be 
read directly off the lexical entries for individual 
verbs, as well-as composed representations for sen- 
tences, using uniform processes and representations. 
The LCS framework consists of primitives (GO, 
BE, STAY, etc.), types (Event, State, Path, etc.) 
and fields (Loc(ational), Temp(oral), Foss(essional), 
Ident(ificational), Perc(eptual), etc.). 
We adopt a refinement of the LCS representation, 
incorporating meaning components from the linguis- 
tically motivated notion of !ezical semantic template 
(LST), based on lexical aspect classes, as defined 
in the work of Levin and Rappaport Hovav (Levin 
and Rappaport Hovav, 1995; Rappaport lttovav and 
Levin, 1995). Verbs that appear in multiple as- 
pectual frames appear in multiple pairings between 
constants (representing the idiosyncratic meaning 
of the verb) and structures (the aspectual class). 
Since the aspectual templates may be realized in 
a variety of ways, other aspects of the structural 
meaning contribute to differentiating the verbs from 
each other. Our current database contains some 400 
classes, based on an initial representation of the 213 
classes in (Levin, 1993). Our current working lexi- 
con includes about 10,000 English verbs and 18,000 
Chinese verbs spread out into these classes. 
Telic verbs (and sentences) contain certain types 
of Paths, or a constant, represented by ! !, filled by 
the verb constant, in the right most leaf-node argu- 
ment. Some examples are shown below: 
depart (go foe (* thing 2) 
(away_from loc (thing 2) 
(at foe (thing 2) 
(* thing 4))) 
(!!+ingly 26)) 
insert (cause (* thing 1) 
(go loc (* thing 2) 
((* toward 5) loc (thing 2) 
(\[at\] loc (thing 2) 
(thing 6)))) 
(! !+ingly 26)) 
Each of these relic verbs has a potential coun- 
terpart with an atelic verb plus the requisite path. 
Depart, for example, corresponds to move away, or 
something similar in another . 
We therefore identify telic sentences by the algo- 
rithm, formally specified in in Figure 1 (cf. (Dorr 
and Olsen, 1997b) \[156\]). 
Given an LCS representation L: 
1. Initialize: T(L):=\[¢T\], D(L):=\[0R\], R(L):=\[0D\] 
2. If Top node of L E {CAUSE, LET, GO} 
Then T(L):=\[+T\] 
If Top node of L E {CAUSE, LET} 
Then D(L):=\[+D\], R(L):=\[+R\] 
If Top node of L E {GO} 
Then D(L):=\[+D\] 
3. If Top node of L E {ACT, BE, STAY} 
Then If Internal node of 
L E {TO, TOWARD, FORTemp} 
Then T(L):=\[+T\] 
If Top node of L E {BE, STAY} 
Then R(L):=\[+R\] 
If Top node of L E {ACT} 
Then set D(L):=\[+D\], R(L):=\[+R\] 
4. Return T(L), D(L), R(L). 
Figure 1: Algorithm for Aspectual Feature Determi- 
nation 
This algorithm applies to the structural primitives 
of the interlingua structure rather than actual verbs 
in source or target . The first step initial- 
ized the aspectual values as unspecified: atelic f-T\], 
stative (not event: f-D\]), and adurative f-R\]. First 
the top node is examined for primitives that indicate 
telicity: if the top node is CAUSE, LET, GO, telicity 
is set to \[+T\], as with the verbs break, destroy, for 
example. (The node is further checked for dynamic- 
ity \[+D\] and durativity \[+R\] indicators, not in focus 
in this paper.)If the top node is not a relic indicator 
(i.e., the verb is a basically atelic predicate such as 
love or run, telicity may still be still be indicated 
by the presence of complement nodes of particular 
types: e.g. a goal phrase (to primitive) in the case of 
run. The same algorithm may be used to determine 
felicity in either individual verbal entries (break but 
37 
not run) or composed sentences (John ran to ~he 
store but not John ran. 
Similar mismatches of telicity between represen- 
tations of particular predicates can occur between 
s, although there is remarkable agreement 
as to the set of templates that verbs with related 
meanings will fit into (Olsen et al., 1998). In the 
Chinese-English interlingual system we describe, the 
Chinese is first mapped into the LCS, a - 
independent representation, from which the target- 
 sentence is generated. Since telicity (and 
other aspects of event structure) are uniformly rep- 
resented at the lexical and the sentential level, telic- 
ity mismatches between verbs of different s 
may then be compensated for by combining verbs 
with other .components. 
.o 
4 Predictions 
Based on (Dowty, 1986) and others, as discussed 
above, we predict that sentences that have a telic 
LCS will better translate into English as the past 
tense, and those that lack telic identifiers will trans- 
late as present tense. Moreover, we predict that 
verbs in the main clause that are telic, will be past 
with respect to their subordinates (X then Y). Verbs 
in the main clause that are atelic we predict will tem- 
porally overlap (X while Y). 
5 Implementation 
LCSes are used as the interlingua for our machine 
translation efforts. Following the principles in (Dorr, 
1993), lexical information and constraints on well- 
formed LCSes are used to compose an LCS for a 
complete sentence from a sentence parse in a source 
. This composed LCS (CLCS) is then used 
as the starting points for generation into the target 
, using lexical information and constraints 
for the target . 
The generation component consists of the follow- 
ing subcomponents: 
Decomposition and lexlcal selection First, 
primitive LCSes for words in the target lan- 
guage are matched against CLCSes, and tree 
structures of covering words are selected. Am- 
biguity in the input and analysis represented 
in the CLCS is maintained (insofar as it is 
possible to realize particular readings using the 
target  lexicon), and new ambiguities 
are introduced when there are different ways of 
realizing a CLCS in the target . 
AMR Construction This tree structure is then 
translated into a representation using the Aug- 
mented Meaning Representation (AMR) syntax 
• of instances and hierarchical relations (Langk- 
fide and Knight, 1998a); however the rela- 
tions include information present in the CLCS 
and LCSes for target  words, including 
theta roles, LCS type, and associated features. 
Realization The AMR structure is then linearized, 
as described in (Dorr et al., 1998), and mor- 
phological realization is performed. The result 
is a lattice of possible realizations, represent- 
ing both the preserved ambiguity from previous 
processing phases and multiple ways of lineariz- 
ing the sentence. 
Extraction The final stage uses a statistical bi- 
gram extractor to pick an approximation of the 
most fluentrealization (Langkilde and Knight, 
1998b). 
While there are several possible ways to address 
the tense and discourse connective issues mentioned 
above, such as modifying the LCS primitive elements 
and/or the composition of the LCS from the source 
, we instead have been experimenting for 
the moment with solutions implemented within the 
generation component. The only extensions to the 
LCS  have been loosening of the constraint 
against direct modification of states and events by 
other states and events (thus allowing composed LC- 
Ses to be formed from Chinese with these structures, 
but creating a challenge for fluent generation into 
English), and a few added features to cover some of 
the discourse markers that are present. We are able 
to calculate telicity of a CLCS, using the algorithm 
in Figure 1 and encode this information as a binary 
teli¢ feature in the Augmented Meaning Represen- 
tation (AMR). 
The realization algorithm has been augmented 
with the rules in (6) 
(6) a. If there is no tense feature, use telicity to 
determine the tense: 
: telic + -~ : tense past 
: relic -- --~ : tense present 
b. In an event or state directly modifying 
another event or state, if there is no other 
clausal connective (coming from a subor- 
dinating conjunction or post-position in 
the original), then use telicity to pick a 
connective expressing assumed temporal 
relation: 
: relic -~ -~ : sconj then 
: relic -- -~ : sconj while 
6 The Corpus 
We have applied this machine translation system to 
a corpus of Chinese newspaper text from Xinhua and 
other sources, primarily in the economics domain. 
The genre is roughly comparable to the American 
38 
Wall Street Journal. Chinese newspaper genre dif- 
fers from other Chinese textual sources, in a number 
of ways, including: 
• more complex sentence structure 
• more extensive use of acronyms 
• less use of Classical Chinese 
• more representative grammar 
• more constrained vocabulary (limited lexicon) 
• abbreviations are used extensively in Chinese 
newspaper headlines 
However, the presence of multiple events and 
states in a single sentence, without explicit modifi- 
catioia is characteristic of written Chinese in general. 
In the 80-sentence corpus under consideration, the 
sentence structure is complex and stylized; with an 
average of 20 words per sentence. Many sentences, 
such as (1)and (2), have multiple clauses that are 
not in a direct complement relationship or indicated 
with explicit connective words. 
7 Ground Truth 
To evaluate the extent to which our Predictions re- 
sult in an improvement in translation, we have used 
a database of human translations of the sentences 
in our corpus as the ground truth, or gold standard. 
One of the translators is included among our au- 
thors. 
The ground truth data was created to provide a 
fluid human translation of the text early in our sys- 
tem development. It therefore includes many com- 
plex tenses and multiple sentences combined, both 
currently beyond the state of our system. Thus, 
two of the authors and an additional researcher 
also created a database of temporal relations among 
the clauses in the sentences that produced illegal 
event/state modifications. This was used to test pre- 
dictions of temporal relationships indicated by telic- 
ity. In evaluating our results, we concentrate on how 
well the System did at matching past and present, 
and on the appropriateness of temporal connectives 
generated. 
8 Results 
We have applied the rules in (6) in generating 80 sen- 
tences in the corpus (starting from often ambiguous 
CLCS analyses). Evaluation is still tricky, since, in 
many cases, the interlingua analysis is incorrect or 
ambiguous in ways that affect the appropriateness 
of the generated translation. 
8.1 Tense 
As mentioned above, evaluation can be very diffi- 
cult in a number of cases. Concerning tense, our 
"gold standard" is the set of human translations, 
generated tense 
past present 
human past 134 17 
translation present 17 27 
Table 2: Preliminary Tense Results 
previously constructed for these sentences. In many 
cases, there is nothing overt in the sentence which 
would specify tense, so a mismatch might not actu- 
ally be "wrong". Also, there are a number of sen- 
tences which were not directly applicable for com- 
parison, such as when the human translator chose 
a different syntactic structure or a complex tense. 
The newspaper articles were divided into 80 sen- 
tences. Since some of these sentences were conjunc- 
tions, this yielded 99 tensed main verbs. These verbs 
either appeared in simple present, past, present or 
past perfect('has or had verb-t-ed), present or past 
imperfective (is verb-l-lag , was verb--I--lag) and their 
corresponding passive (is being kicked, was being 
kicked, have been kicked) forms. For cases like the 
present perfect ('has kicked), we noted the intended 
meaning ( e.g past activity) expressed by the verb 
as well as the verb's actual present perfective form. 
We scored the form as correct if the system trans- 
lated a present perfective with past tense meaning 
as a simple past or present perfective. There were 
10 instances where a verb in the human translation 
had no corresponding verb in the machine transla- 
tion, either due to incorrect omission or correct sub- 
stitution of the corresponding nominalization. We 
excluded these forms from consideration. If the sys- 
tem fails to supply a verb for independent reasons, 
our system clearly can't mark it with tense. The 
results of our evaluation are summarized in Table 2. 
These results definitely improve over our previ- 
ous heuristic, which was to always use past tense 
(assuming this to be the default mode for newspa- 
per article reporting). Results are also better than 
always picking present tense. These results seem to 
indicate that atelicity is a fairly good cue for present 
tense. We also note that 8 out of the 14 cases where 
the human translation used the present tense while 
the system used past tense are headlines. Headlines 
are written using the historical present in English 
("Man bites Dog"). These sentences would not be 
incorrectly translated in the past ("The Man Bit 
the Dog") Therefore, a fairer judgement might leave 
only remaining 6 incorrect cases in this cell. Using 
atelicity as a cue for the present yields correct re- 
sults approximately 65incorrect results 35worst case 
results because they do not take into account pres- 
ence or absence of the grammatical perfective and 
progressive markers referred to in the introduction. 
39 
8.2 Relationship between clauses 
Results are more preliminary for the clausal connec- 
tives. Of the 80 sentences, 35 of them are flagged as 
(possibly) containing events or states directly mod- 
ifying other events or states. However, of this num- 
ber, some actually do have lexical connectives repre- 
sented as featural rather than structural elements in 
the LCS, and can be straightforwardly realized using 
translated English connectives such as since, after, 
and if.then. Other apparently "modifying" events 
or states should be treated as a complement rela- 
tionship (at least according to the preferred reading 
in ambiguous cases), but are incorrectly analyzed 
as being in a non-complement relationship, or have 
other structural problems rendering the interlingua 
representation and English output not directly re- 
lated to the original clause structure. 
Of the remaining clear cases, six while relation- 
ships were generated according to our heuristics, in- 
dicating cotemporaneousness of main and modifying 
situation, e.g. (7)a,b, in the automated translations 
of (1) and (2), respectively. None were inappropri- 
ate. Of the cases where then was generated, indicat- 
ing sequential events, there were four cases in which 
this was appropriate, and three cases in which the 
situations really should have been cotemporaneous. 
While these numbers are small, this preliminary data 
seems to suggest again that atelicity is a good cue for 
cotemporality, while telicity is not a sufficient cue. 
(7) a. Before 1965, China altogether only have 
the ability shipbuilding about 300 thousand 
tons , while the annual output is 80 thou- 
sand tons. 
b. this 80 thousand tons actually includes 517 
ships, while the ship tonnage is very low. 
9 Conclusions 
We therefore conclude that lexical aspect can serve 
as a valuable heuristic for suggesting tense, in the ab- 
sence of tense and other temporal markers. We an- 
ticipate incorporation of grammatical aspect infor- 
mation to improve our temporal representation fur- 
ther. In addition, lexical aspect, as represented by 
the interlingual LCS structure, can serve as the foun- 
dation for  specific heuristics. Furthermore, 
the lexical aspect represented in the LCS can help to 
provide the beginnings of cross-sentential discourse 
information. We have suggested applications in the 
temporal domain while, then. Causality is another 
possible domain in which relevant pieces encoded in 
sentence-level LCS structures could be used to pro- 
vide links between LCSes/sentences. Thus, the in- 
terlingual representation may be used to provide not 
only shared semantic and syntactic structure, but 
"also the building blocks for -specific heuris- 
tics for mismatches between s. 
10 Future Research 
There are a number of other directions we intend to 
pursue in extending this work. First, we will evalu- 
ate the role of the grammatical aspect markers men- 
tioned above, in combination with the telicity fea- 
tures. Second, we will also examine the role of the 
nature of the modifying situation. Third, we will 
incorporate other lexical information present in the 
sentence, including adverbial cue words (e.g. now, 
already and specific dates that have time-related in- 
formation, and distinguishing reported speech from 
other sentences. Finally, as mentioned, these re- 
sults do not take embedded verbs or verbs in adjunct 
clauses into account. Many adjunct and embedded 
clauses are tenseless, making evaluation more diffi- 
cult. For example, is The President believed China 
to be a threat equivalent to The president believed 
China is a threat). 

References 
Chauncey C. Chu. 1998. A Discourse Grammar of 
Mandarin Chinese. Peter Lang Publishing, Inc., 
New York, NY. 
Bernard Comrie. 1976. Aspect. Cambridge Univer- 
sity Press, Cambridge, MA. 
Bonnie J. Dorr and Terry Gaasterland. 1995. Se- 
lecting Tense, Aspect, and Connecting Words in 
Language Generation. In Proceedings of IJCAI- 
95, Montreal, Canada. 
Bonnie J. Dorr and Marl Broman Olsen. 1997a. As- 
pectual Modifications to a LCS Database for NLP 
Applications. Technical Report LAMP TR 007, 
UMIACS TR 97-23, CS TR 3763, University of 
Maryland, College Park, MD. 
Bonnie J. Dorr and Marl Broman Olsen. 1997b. 
Deriving Verbal and Compositional Lexical As- 
pect for NLP Applications. In Proceedings of the 
35th Annual Meeting of the Association for Com- 
putational Linguistics (ACL-97), pages 151-158, 
Madrid, SPain , July 7-12. 
Bonnie J. Doff, James Hendler, Scott Blanksteen, 
and Barrie Migdaloff. 1993. Use of Lexical Con- 
ceptual Structure for Intelligent Tutoring. Tech- 
nical Report UMIACS TR 93-108, CS TR 3161, 
University of Maryland. 
Bonnie J. Dorr, Nizar Habash, and David Traum. 
1998. A Thematic HieJfarchy for Efficient Gener- 
ation from Lexical-ConceptM Structure. In Pro- 
ceedings of the Third Conference of the Associ- 
ation for Machine Translation in the Americas, 
AMTA-98, in Lecture Notes in Artificial Intelli- 
gence, 15~9, pages 333-343, Langhorne, PA, Oc- 
tober 28-31. 
Bonnie J. Dorr. 1993. Machine Translation: A View 
from the Lexicon. The MIT Press, Cambridge, 
MA. 
David Dowty. 1979. Word Meaning in Montague 
Grammar. Reidel, Dordrecht. 
David Dowty. 1986. The Effects of Aspectual Class 
on the Temporal Structure of Discourse: Seman- 
tics or Pragmatics? Linguistics and Philosophy, 
9:37-61. 
Mohamed Guerssel, Kenneth Hale, Mary Laugh- 
ten, Beth Levin, and Josie White Eagle. 1985. 
A Cross-linguistic Study of Transitivity Alterna- 
tions. In W. H. Eilfort, P. D. Kroeber, and K. L. 
Peterson, editors, Papers from the Parasession on 
Causatives and Agentivity at the Twenty.First Re- 
gional Meeting, CLS P1, Part P, pages 48-63. The 
Chicago Linguistic Society, Chicago, IL, April. 
Ray Jackendoff. 1983. Semantics and Cognition. 
The MIT P.r¢ss, Cambridge, MA. 
Ray Jackendoff. 1990.. Semantic Structures. The 
MIT Press, Cambridge, MA. 
Irene Langkilde and Kevin Knight. 1998a. Gen- 
eratiort that Exploits Corpus-Based Statistical 
Knowledge. In Proceedings of COLING-ACL '98, 
pages 704-710. 
Irene Langkilde and Kevin Knight. 1998b. The 
Practical Value of N-Grams in Generation. In In- 
ternational Natural Language Generation Work- 
shop. 
Beth Levin and Malka Rappaport Hovav. 1995. Un- 
accusativity: At the Syntaz-Lexical Semantics In- 
terface. The MIT Press, Cambridge, MA. LI 
Monograph 26. 
Beth Levin. 1993. English Verb Classes and Alter- 
nations: A Preliminary Investigation. University 
of Chicago Press, Chicago, IL. 
Charles Li and Sandra Thompson. 1981. Mandarin 
Chinese: A functional reference grammar. Uni- 
versity of California Press, Berkeley, CA. 
Marc Moens and Mark Steedman. 1988. Tempo- 
ral Ontology and Temporal Reference. Compu- 
lational Linguistics: Special Issue on Tense and 
Aspect, 14(2):15-28. 
Mart Broman Olsen, Bonnie J. Dorr, and Scott C. 
Thomas. 1998. Enhancing Automatic Acquisi- 
tion of Thematic Structure in a Large-Scale Lex- 
icon for Mandarin Chinese. In Proceedings of the 
Third Conference of the Association for Machine 
Translation in the Americas, AMTA-98, in Lec- 
ture Notes in Artificial Intelligence, 1529, pages 
41-50, Langhorne, PA, October 28-31. 
Mart Broman Olsen. 1997. A Semantic and Prag- 
matic Model of Lexical and Grammatical Aspect. 
Garland, New York. 
Rebecca Passoneau. 1988. A Computational Model 
of the Semantics of Tense and Aspect. Compu- 
tational Linguistics: Special Issue on Tense and 
Aspect, 14(2):44-60. 
Malka P~appaport Hovav and Beth Levin. 1995. 
The Elasticity of Verb Meaning. In Processes in 
Argument Structure, pages 1-13, Germany. SfS- 
Report-06-95, Seminar fiir Sprachwissenschaft, 
Eberhard-Karls-Universit~t Tiibingen, Tiibingen. 
Carlota Smith. 1997. The parameter of aspect. 
Kluwer, Dordrecht, 2nd edition. 
