APPROXIMATING AN INTERLINGUA 
IN A PRINCIPLED WAY 
Eduard Hovy I and Sergei Nirenburg 2 
1) Information Sciences Institute 2) Center for Machine Translation 
4676 Admiralty Way Carnegie Mellon University 
Marina del Rey, CA 90292-6695 Pittsburgh, PA 15213 
ABSTRACT 
We address the problem of constructing in a principled way 
an ontology of terms to be used in an interlingua for ma- 
chine translation. Given our belief that the a true language- 
neutral ontology of terms can only be approached asymp- 
totically, the construction method outlined involves a step- 
wise folding in of one language at a time. This is effected 
in three steps: first building for each language a taxonomy 
of the linguistic generalizations required to analyze and gen- 
erate that language, then organizing the domain entities in 
terms of that taxonomy, and finally merging the result with 
the existing interlingua ontology in a well-defined way. This 
methodology is based not on intuitive grounds about what 
is and is not 'true' about the world, which is a question of 
language-independence, but instead on practical concerns, 
namely what information the analysis and generation pro- 
grams require in order to perform their tasks, a question of 
language-neutrality. After each merging is complete, the re- 
sulting taxonomy contains, declaratively and explicitly rep- 
resented, those distinctions required to control the analysis 
and generation of the linguistic phenomena. The paper is 
based on current work of the PANGLOSS MT project. 
1. Interlinguas 
This paper presents a method of constructing in a prin- 
cipled way an ontology of terms to be used as a source 
of terms in an interlingua for machine translation. The 
method involves taxonomizing relevant linguistic phe- 
nomena in each language and merging the resulting tax- 
onomy with the interlingua ontology to produce an on- 
tology that explicitly records the phenomena that must 
be handled by any parser or generator and is neutral 
with respect to the languages handled by the system. 
1.1. What is an Interlingua? 
In interlingual machine translation, the representational 
power of the interlingua is central to the success of the 
translation. By interlingua we mean a notation used 
in MT systems to represent the propositional and prag- 
matic meaning of input texts; an interlingua text that 
represents the meaning of a source language text is pro- 
duced by computational analysis, and is then used as 
input to a generation module which realizes this mean- 
ing in a target language text. 
An interlingua consists of the following three parts: 
a collection of terms: the elements that represent 
individual meanings (of lexical items, pragmatic as- 
pects, etc.); this collection is organized in a multiply 
interconnected semantic network; 
notation: the syntax to which well-formed interlin- 
gua texts conform; 
substrate: the knowledge representation system in 
which interlingua texts are instantiated and which 
provided the infrastructure for reasoning with the 
knowledge encoded in interlingua texts. 
The term network can be thought of as a conceptual lex- 
icon of basic meaning distinctions that represent entities 
in the world. While it is not feasible to produce a com- 
plete model of the world, it is possible to produce a well- 
formed subset in which each term's definition contains 
enough features to differentiate it from all other terms 
in the subworld. A good taxonomization enables the 
sharing and inheritance of properties and facilitates def- 
initional brevity and expressive power. Any given self- 
contained subworld includes a set of terms which it can 
share with other subworlds. These are to be found at 
higher levels of the network, closer to its root node. It 
is customary to call a subnetwork of terms that are ex- 
pected to be present in most (if not all) subworlds an 
ontology, while the rest of a particular subworld is often 
called a domain model. Though the boundary between 
an ontology and a set of domain models is rather diffi- 
cult to establish in the abstract, in practice, for any given 
set of domain models, it is straightforward to detect the 
shared higher levels. 
The main purpose of an ontology and a domain model 
in MT is to support lexical-semantic analysis and gen- 
eration. The open-class lexical units of a language have 
their meanings explained using terms from the ontology 
and the domain model. Connections are made in the 
lexicons for each particular language. (For a detailed 
description of one such interlingua-ontology-lexicon com- 
bination see \[15, 13, 4\]). 
261 
1.2. Language Independence vs. 
Language Neutrality 
By definition, an interlingua is language-independent 
and depicts the world directly, not through the medium 
of a natural language. Of course, particular interlinguas 
vary in their quality and utility. The two major proper- 
ties of an interlingua which influence its quality are its 
expressive power (that is, the ability to support a dif- 
ferent structure for each of a set of ambiguous natural 
language expressions) and its capability to scale up with- 
out a major overhaul of its syntax (a point which relates 
directly to the task of term acquisition). 
In theory, interlingua is language-independent. That is, 
all the languages handled are pairwise independent with 
respect to the interlingua, and the interlingua contains 
no "traces" of particular semantic distinctions predomi- 
nant in a particular language or group of languages. The 
degree of language independence attained by an interlin- 
gua in reality is a matter of judgment. 
The feasibility of language-independent representations 
has been debated many times. Since interlinguas are 
by definition language-independent, the debate revolves 
about the degree of overlap with language-specific knowl- 
edge and the status of such overlap. This debate is inter- 
esting but not practically relevant in the area of MT. A 
fully language-independent ontology is a worthy ideal. 
Since the acquisition-related tasks of selection, defini- 
tion, and taxonomization of terms for an ontology and 
domain model are always performed by people, they are 
always open to being influenced by the distinctions made 
in the acquirers' languages as well as their "folk" models 
of the world. 1 
However, our approach to the problem of interlingua con- 
struction is more practical. We are concerned with the 
matter of constructing an interlingua that is language- 
neutral, i.e., one that enables analyzers and generators 
to operate optimally in a machine translation system. 
Under this approach it is still possible to maintain the 
central architectural tenet of interlingual MT systems -- 
that it is possible to translate from a text to its interlin- 
gua text without regard to any eventual target language, 
and to translate from the interlingua text to a target 
without regard to the original source language. 
We focus in this paper on the ontology, leaving aside 
questions of notation and substrate. Given the complex- 
1 Consider, for example, the problem of representing colors: how 
many are there, and how are they organized? Even the basic dis- 
tinction into the three primary colors so natural to any westerner 
is unlikely to occur to the Daui people of New Guinea, who have 
only two color terms (mill = dark-cool; mola = light-warm) and 
have proven difficulty in learning names for nonfocM colors \[16\]. 
ity of the ontology construction task, we believe that 
any given ontology will inevitably be closer to some lan- 
guages than to others, both in the terms selected and 
their connections. We don't consider this to be a prob- 
lem, as we would be able to augment the ontology and 
domain model as necessary whenever new domain mod- 
els and new languages are added to the system. 
What we do consider a problem is the lack of a generally 
established well-defined method for constructing ontolo- 
gies. Most ontologies and domain models to date have 
been assembled based primarily on introspection, and 
often reflect the idiosyncrasies of the builders more than 
the requirements of the application (such as MT). Lack- 
ing well-founded guiding principles, the ontology builder 
is working in the dark. In this paper, we suggest that 
a better methodology uses data-oriented methods and 
linguistic knowledge to guide introspection. 
2. A Method of Ontology Construction 
2.1. Basic Criterion of Taxonomization 
Recognizing that we cannot build a satisfactorily 
language-neutral ontology directly, we formulate an in- 
cremental procedure that, we believe, allows the acquirer 
to approach an optimal ontology asymptotically. 
To develop the methodology we seek theoretically moti- 
vated answers to the following basic questions underlying 
ontology construction: 
• What terms must be included in the ontology? 
• How should the terms be organized in it? 
• What level of detail should it reach? 
• How closely can the ontology terms parallel the par- 
ticular words and phrases of any single language? 
In answering these questions, we exploit the fact that the 
ontology is inevitably going to be more or less language- 
specific to further our overall goal: a powerful interlin- 
gual MT system. The "closer" the ontology is to the 
source and target languages, the easier the process of 
ontology term acquisition and organization, and the less 
work the parser and generator have to do to bridge the 
gap between interlingua texts and source and target lan- 
guage texts. To find the point of "minimal distance" 
from all source and target languages, weighted as dis- 
cussed in Section 2.5, we formulate the: 
Basic criterion underlying terminology 
creation and taxonomization: The ontology 
must be just sufficiently powerful to represent 
the distinctions required to analyze the source 
languages and generate the target languages in 
the given domains. 
262 
To define the distinctions needed to support each lan- 
guage, we start with a list of linguistic phenomena to be 
covered in it. This information can be extracted from a 
sufficiently rich grammar of the language. For example, 
for English the fact that nouns pattern into mass and 
count, the fact that adjectives and adverbs pattern dif- 
ferently than do verbs, or the fact that many different 
forms of possession (to have an arm, to have a spouse, 
to have a car) are expressed similarly. We then create 
a taxonomy of these linguistic abstractions, guided by 
their interrelationships. This taxonomy would, for these 
examples, contain nodes for UncountableObj and Count- 
ableObj under Obj to help handle nouns, nodes Quality 
and Process as high-level taxonomic organizers of adjec- 
tival/adverbial modifiers and verbal actions, and Gen- 
eralizedPossession as a high-level organizational point 
for the various types of possession. Any proposed term 
is accepted in the taxonomy only if it captures a dis- 
tinct linguistic phenomenon of the language not oth- 
erwise handled. We end up with a "general" taxon- 
omy of abstract entities that capture useful groupings 
of processing-oriented features. 
Having established such a taxonomy, we then list all the 
entities that appear in the domain being addressed by 
the MT system (objects, actions, states, relations, plans, 
scripts (action chains), etc.). What we are after is a tax- 
onomization of these entities in such as way as to facil- 
itate MT. Of course, the simplest scheme is no taxono- 
mization at all: a list of the domain entities without any 
higher-level organizing entities. Then however each en- 
tity must contain enough information by itself to enable 
successful manipulation by the analyzer and generator 
(for example, Company must be explicitly annotated as 
being a SociaiOrg, if this fact is used by analyzer and/or 
generator). A better solution is to group entities with 
the same features together and define nodes that repre- 
sent generalizations. But which generalizations? We ar- 
gue that most useful are those generalizations that play 
a role in the processing, analysis and generation (those 
that serve some differentiating function in the construc- 
tion of an analysis or a clause) -- exactly those contained 
in the taxonomy of linguistic phenomena. This taxon- 
omy partitions the entities of the domain in ways which 
facilitate their treatment by the system. ~ 
Thus, we use the linguistically inspired taxonomy as on- 
2It must be understood that we are not stating that one must 
(or even can) perform exhaustive decomposition of all the facets 
of meaning of a concept to place it in the ontology. The basic 
criterion argues that just enough must be represented to enable 
the generation of the languages in question. Thus for example we 
do not need a term Pink for Pale Red unless the domain is such 
that there are implicative effects to saying "pale red" instead of 
"pink" (the former being a noncentral meaning in some contexts; 
see \[12\] and \[8\], p. 115). 
tology under which to categorize all the domain terms 
(though this is not the only way to employ the language- 
oriented taxonomy as a guide, it is the most straightfor- 
ward). By such a taxonomization, each domain term is 
located so as to inherit precisely the representational dis- 
tinctions required to support its generation in the target 
language. Such taxonomies are fairly common in sym- 
bolically based NLP systems; see for example the ones 
used in JANUS (BBN) \[7\], LILOG (IBM Germany) \[11\], 
and Naive Semantics \[5\]. 
We call the resulting taxonomy, which consists of a lin- 
guistically inspired ontology and a domain-oriented do- 
main model, the Language Base (LB) of the language. 
Since the resulting knowledge base is language-based, 
it is not necessarily best suited to support specialized 
forms of reasoning such as naive physics, legal reasoning, 
etc. But since our intention is MT, this poses no prob- 
lem. An example of the abstractions that can be used as 
high-level organization for English is the Penman Upper 
Model \[2\], a taxonomy that has been used to support the 
generation of English text in several domains. 
2.2. Toward Language Neutrality 
It will be noted that the example terms employed above 
are semantic rather than syntactic (e.g., Process rather 
than Verb). Where obvious semantic correlates for syn- 
tactic generalizations exist, the strategy is to employ 
them, for they are more likely to appear also in the LBs 
of other languages. However, the result is still (at best) 
a mixture of semantic and purely syntactic generaliza- 
tions, including generalizations for phenomena without 
any obvious semantic basis; for example, in English, the 
hct that verbs like "fill" and "spray" pattern similarly: 
"he filled the cart with hay" and "he sprayed the wall 
with paint" can be similarly transformed to "he used hay 
to fill the cart" and "he used paint to spray the wall". 
Using abstractions from any particular language means 
giving up the basic goal of language neutrality. Since, 
however, we expect that any ontology is going to be 
slanted toward some language(s) more than others, we 
are willing to accommodate such "syntactic pollution" 
as long as it does not hamper MT. We next outline 
a method of progressively making a taxonomy more 
language-neutral to the point where all troublesome syn- 
tactic terms are removed. This process involves merging 
the LB for a second language with the LB for the first 
to form a hybrid taxonomy (also called a polylingua in 
\[9\]), which we call the Ontology Base (OB), and then re- 
peating the process for additional languages, according 
to the following procedure: 
1. Construct the LB taxonomy for language 1. This is 
the ontology base (OB). 
263 
2. For each subsequent language, 
(eL) construct the LB for its phenomena; 
(b) merge it with the existing OB, starting from 
the topmost entity and proceeding downward, 
as described in Section 2.3, ensuring that the 
lower-level, domain-specific, OB terms remain 
properly taxonomized. 
The merging process can be considerably simplified if 
during construction of the LB for a new language (step 
2(a)), the classes of the existing OB are used as a guide 
whenever various taxonomizations are possible. 
From the languages we have examined, the following 
trend is apparent: the upper regions (ontologies) of the 
LBs are identical or nearly so. Differences occur in the 
middle regions, since they reflect language-particular in- 
formation; the degree of cross-LB commensurability de- 
pends roughly on the closeness of the grammars. The 
lower regions are essentially identical for fairly inter- 
national domains such as banking, science, technology, 
etc. -- after all, however something like Metal is treated 
in the language, its descendants Gold, Silver, etc., are 
treated similarly and thus taxonomize together similarly. 
As discussed in \[17\] and \[10\], conceptual systems (of 
which OBs and LBs are an example) can be commen- 
surate in several different ways. Lakoff lists five types 
of commensurability of two conceptual systems (p. 322): 
translation (a sentence-by-sentence translation preserv- 
ing truth conditions is possible), understanding (a per- 
son can understand both alternatives), use (the same 
concepts are used the same ways), framing (situations 
are "framed" the same way and there is a one-one cor- 
respondence between the systems, frame by frame), and 
organization (when the same concept occurs in both, it 
is organized the same way relative to others that occur 
in both). Lakoff provides some examples: the systems of 
aspect in English and Hopi are commensurate by transla- 
tion (since Whorf did translate sentences from one to the 
other) but not by use, since as Whorf's examples make 
clear (pp. 57-64), the "same" concepts for time are used 
very differently; the systems of Spatial Location in En- 
glish and Mixtec are commensurate by translation but 
not by organization \[3\], since for example sentences ex- 
pressing the English meaning "on" translate to widely 
different Mixtec expressions depending on the shape of 
the lower object, among other things. 
2.3. Merging Ontologies 
Given the OB and an LB for a new language, we start 
the merging process from the topmost item(s) of the hi- 
erarchies. Usually, since the "meaning" of each item 
is captured by its interrelationships and ancestry, it is 
instructive to consider groups of closely related items si- 
multaneously. The merging process involves one of three 
alternative operations for each item in the LB: 
Identity: the LB item is identical to a corresponding 
item of the OB; they represent the same phenomenon, 
such as DecomposableObj. In this case, no further work 
is required beyond a name change to the OB item name. 
Identity may be difficult to determine for non-semantic 
items (one reason we suggest using semantic items when 
possible), or even occasionally for semantic ones, when 
their subordinate items differ. We discuss this point later 
in this section. However, in practise, identity of OB and 
LB items is common for related languages, especially in 
the more abstract (higher) regions and more domain- 
specific (lower) regions of the OB. 
Extension: the LB item is more specific than the ap- 
propriate OB item(s); that is, it straightforwardly sub- 
categorizes some OB item(s). In this case, the OB is 
grown by including the LB item as a child of the OB 
item. For example, if the OB were initially constructed 
from English, its system of honorifics would probably 
contain only two items, one for FormalSuperordinate and 
the other for EqualOrSubordinate; a Japanese LB would 
cause this system to be fleshed out to include a more 
elaborate substructure along the lines described in \[1\]. 
Cross-elasslfication: the LB item represents aspects 
of more than one OB items. This is the case the new 
language partitions the phenomenon under considera- 
tion in a different way to the previous language(s) stud- 
ied. Typically, several parallel LB items represent one 
partitioning of the phenomenon and several OB items 
represent a different partitioning. In this case, two al- 
ternative strategies can be followed. The first is to enter 
the LB items into the OB as a parallel but distinct tax- 
onomization of the phenomenon, and all their descen- 
dants must be added as well, unless items representing 
the same descendants are already in the OB, in which 
case these items must be linked up also to the appro- 
priate LB item(s). The second strategy is to create the 
cross-product of the two sets of items. For example, if 
the OB partitions Objs into UncountableObj (mass) and 
CountableObj (count) types, and the new language parti- 
tions Objs into (say) TaiiSkinnyObj and OtherObj types, 
and neither LB class is a proper subset of either OB 
one, then four new items must be formed: Countable- 
TallSkinnyObj, Uncountable-TallSkinnyObj, Countable- 
OtherObj, and Uncountable-OtherObj. Every item sub- 
ordinate to either item in both LB and OB must then be 
reclassified with respect to the new cross-product items. 
One difficult case arises when the same item is tax- 
onomized in the LB and OB under mutually exclusive 
264 
items. For example, the domain concept Sand may be 
classified as UncountableObj in one and CountableObj 
in the other. In this case, the respective LB and OB 
items UncountableObj and CountableObj must be differ- 
entiated as, for example, UncountableObjl and Count- 
ableObjl and UncountableObj2 and CountableObj2, the 
cross-product taken, and any subordinate items reclas- 
sifted. Though this may become a combinatorially ex- 
pensive operation in principle, in practise we find it sel- 
dom occurs in our domain; furthermore, the fact that all 
three our current LBs for English contain fewer than 200 
non-domain items (see Section 3), and we do not expect 
significantly larger LBs to be necessary. 
2.4. Functional Grounds for Ontology 
A certain amount of leeway exists in the construction 
of the LB. This leeway should be used wisely, since, as 
mentioned above, similarity with the existing OB can 
significantly simplify the merging process. For best re- 
sults, linguistic phenomena should be modeled accord- 
ing to their underlying functional reasons. Functional 
distinctions often have different syntactic expressions in 
different languages; it is therefore important to repre- 
sent in the interlingua the function and not the form 
(for example, the rule "passive translates to passive" ig- 
nores the reason why passive was used in the original and 
cannot ensure an appropriate target expression). Un- 
derstanding the true source of variations and creating in 
the ontology appropriate means for representing them is 
of central importance. 
2.5. Tradeoffs in Ontology Content 
The ontology construction method outlined here pro- 
duces a set of terms explicitly defined to represent all 
the pertinent linguistic phenomena of each language be- 
ing covered. For practical purpose, however, this proce- 
dure may introduce more complexity than savings in the 
case where some phenomenon carries meaning only in 
one of the languages being handled. If for example only 
one language differentiated Number into Single, Dual, 
and Multiple (as Arabic and Hebrew do) while all the 
others just differentiated Single from Multiple, the Dual 
option can be removed and handled only by the Arabic 
and Hebrew generators, who in the case of Multiple have 
to determine (somehow) whether the entity in question 
is dual or not, or whether they should simply gener- 
ate an alternative locution. For any particular language, 
the more specifically attuned the LB upper regions are 
to specific forms of expression, the more syntactically 
oriented information enters the ontology. As a result, 
though LB construction for that language is simplified, 
its merging with the OB is made more complex, as is sub- 
sequent merging of the OB with other LBs. In addition, 
analysis of other languages is complicated, since infor- 
mation must be sought that may not be present in the 
source text. To top it all, this information is totally ir- 
relevant when translation occurs to languages other than 
the complexifying culprit. 
Easily recognized during step 2(b) of the merging pro- 
cess of LB and OB by the lack of comparable items in the 
OB, such unique LB phenomena can be omitted from the 
Interlingua, ignored during analysis, and incorporated 
during a pre-realization phase in generation, either by 
using default values, preselected settings, or by turning 
to a human (in-editor or post-editor) for help. The rela- 
tive costs of incorporating such phenomena into the LB 
versus leaving them out and handling them when needed 
depend on the following: 
• the complexity of the phenomenon (which is pro- 
portional to the number of LB items required to 
represent it), 
• ease of handling it by default or circumvention, 
• the frequency of translation into the language(s) ex- 
hibiting the phenomena. 
When LB items are swallowed into specific language pro- 
cessors, a record of the removal of the phenomena should 
be left in the OB at the appropriate point(s) in the tax- 
onomy. Such a record will be encountered during later 
addition of other languages, enabling arguments to be 
made for the reintroduction of the phenomenon into the 
OB itself if appropriate. 
In this manner, the Interlingua may be more or less 
language-neutral, requiring correspondingly more work 
to generate the more distant ones. Its optimal position- 
ing requires a careful analysis of the tradeoffs involved. 
3. Current Work 
We recently began constructing an Interlingua for the 
PANGLOSS machine translation system, using the termi- 
nologies developed by the three partners, namely a set 
of ontologies and domain models developed at the Cen- 
ter for Machine Translation at CMU, using the ONTOS 
acquisition interface \[14\], IR, the Intermediate Repre- 
sentation terminology used at the Computing Research 
Laboratory of New Mexico State University \[6\], and the 
Upper Model developed for the Penman language gen- 
erator at USC/ISI \[2\]. Both CMT's TAMERLAN and 
IR have already been used to support Interlingual ma- 
chine translation of several languages, while variants of 
the Upper Model suited for German, Japanese, and Chi- 
nese are under construction at GMD/IPSI (Germany) 
and the University of Sydney (Australia). 
265 
As expected, except for names, we found little disagree- 
ment among the three ontologies in their upper regions 
(while the IR is not explicitly taxonomized into an on- 
tology, this can be done without problem). The most 
recent ontology from CMU contains approximately 185 
items, IR approximately 150, and the Penman Upper 
Model approximately 190 (the latter two contain only 
linguistic generalizations, no domain entities). 
4. Conclusion 
In this paper we addressed the problem of constructing 
in a principled way one of the three components of an In- 
terlingua for machine translation: the ontology of terms 
which are used in the Interlingua notation to capture the 
meaning of the source text and govern the generation of 
the target text. The methodology described is based 
not on intuitive grounds about what is and is not 'true' 
about the world, which is a matter for philosophers and 
concerns the question of language-independence, but is 
based instead on more practical concerns, namely what 
information the analysis and generation programs re- 
quire in order to perform their tasks. 
Given our belief that the a true language-neutral on- 
tology of terms can only be approached asymptotically, 
the method we outline for constructing the ontology in- 
volves a stepwise folding in of one language at a time. 
Any "syntactic pollution" present in an LB will either be 
merged into the ontology and stand explicitly as some- 
thing other language analyzers and generators have to 
handle or it will be swallowed in the analyzer and gener- 
ator for the specific language as a clearly identified item 
that requires special treatment. In either case, an ex- 
plicit and declarative representation of the distinctions 
required to control the analysis and generation of the 
languages handled results. 
Though (depending on the nature of the analysis and 
generation procedures) a taxonomy of this kind need not 
always resemble what most people intuitively think of 
when they talk about an interlingua ontology, we claim 
this is a moot point, since what we and they are after is 
a practical construct that can be used effectively in an 
MT system, and the method of construction outlined in 
this paper is a way of achieving one. 
References 
1. Bateman, J.A. 1988. Aspects of Clause Politeness in 
Japanese: An Extended Inquiry Semantics Treatment. 
In Proceedings of the ~6th Annual Conference of the 
ACL, Buffalo (147-154). 
2. Bateman, J.A., Kasper, R.T., Moore, J.D. and Whit- 
ney, R.A. 1989. A General Organization of Knowledge 
for Natural Language Processing: The Penman Upper 
Model. Unpublished research report, USC/Information 
Sciences Institute, Marina del Rey. 
3. Brugman, C. 1983. The Use of Bady-Part Terms as 
Locatives in Chalcatongo Mixtec. In Report no. 4 of 
the Survey of California and Other Indian Languages, 
University of California at Berkeley (235-290). 
4. Carlson, L. and Nirenburg, S. 1990. World Modeling for 
NLP. Technical Report no. CMU-CMT-90-121, Carnegie 
Mellon University, Pittsburgh. 
5. Dahlgren, K., McDowell, J., and Stabler, E.P. 1990. 
Knowledge Representation for Commonsense Reasoning 
with Text. Computational Linguistics 15 (149-170). 
6. Farwell, D. 1990. Description of the Intermedia Rep- 
resentation System for the CRL Multilingual Machine 
Translation System. Unpublished document, Comput- 
ing Research Laboratory, New Mexico State University, 
Las Cruces. 
7. Hinrichs, E.W., Ayuso, D.M., Scha, R. 1987. The Syn- 
tax and Semantics of the JANUS Semantic Interpretation 
Language. In Research and Development in Natural Lan- 
guage Understanding as Part o\] the Strategic Computing 
Program. Annual Technical Report no. 6552, BBN Lab- 
oratories, Cambridge. 
8. Jackendoff, R. 1985. Semantics and Cognition. Cam- 
bridge: MIT Press. 
9. Kay, M., Gawron, J.M., and Norvig, P. 1991. Verbmo- 
bil: A Translation System for Face-to-Face Dialog. CSLI 
Study. 
10. Lakoff, G. 1987. Women, Fire, and Dangerous Things. 
Chicago: University of Chicago Press. 
11. Lang, E. 1991. The LmoG Ontology from a Linguistic 
Point of View. In O. Herzog and C-R. Rollinger (eds), 
Text Understanding in LILOG. Berlin: Springer (464- 
480). 
12. McCawley, J.D. 1978. Conversation Implicature and the 
Lexicon. In P. Cole, ed. Syntax and Semantics vol. 9. 
New York: Academic Press (245-259). 
13. Meyer, I., Onyshkevych, B., and Carlson, L. 1990. Lex- 
icographic Principles and Design for Knowledge-Based 
Machine Translation. Technical Report no. CMU-CMT- 
90-118. Carnegie Mellon University, Pittsburgh. 
14. Monarch, I. 1989. ONTOS Reference Manual. Technical 
Memo, Center for Machine Translation. Carnegie Mellon 
University, Pittsburgh. 
15. Nirenburg, S. and Defrise, C. 1992. Application- 
Oriented Computational Semantics. In R. Johnson and 
M. Rosner (eds.), Computational Linguistics and Formal 
Semantics. Cambridge: Cambridge University Press. 
16. Rosch, E. 1973. Natural Categories. Cognitive Psychol- 
ogy 4 (328-350). 
17. Whorf, B.L. 1956. Language, Thought, and Reality: Se- 
lected writings o/ Benjamin Lee Whorl, ed. John B. Car- 
roll, Cambridge: MIT Press. 
266 
