 
Sergei Nirenburg Marjorie McShane Stephen Beale  
 
Institute for Language Information Technologies 
University of Maryland, Baltimore County 
{sergei,marge,sbeale}@umbc.edu 
 
 
Abstract 
 
In this paper, we briefly and informally illus-
trate, using a few annotated examples, the 
static and dynamic knowledge resources of on-
tological semantics. We then present the main 
motivations and desiderata of our approach 
and then discuss issues related to making onto-
logical-semantic applications feasible through 
the judicious stepwise enhancement of static 
and dynamic knowledge sources while at all 
times maintaining a working system.  
 
1.  Introduction 
 
This paper discusses selected issues in ontological se-
mantics (OS), an implemented computational-semantic 
theory that deals with the extraction, representation and 
use of meaning in natural language texts. Unlike practi-
cally all other work in computational semantics, OS 
makes itself responsible for all the necessary compo-
nents and stages in automatic text meaning analysis: it 
addresses lexical and compositional meaning as well as 
pragmatics and discourse issues. Its processing heuris-
tics are derived from syntax, morphology and other 
“preprocessing,” non-semantic analysis stages that are 
still incorporated in the system, as well as from detailed 
underlying world models that include specifications not 
only of basic events, objects and properties but also of 
complex events, or scripts.  
The goal of OS is the extraction, representation and 
manipulation of meaning in natural language texts with 
a view toward supporting applications such as MT or 
question answering. Text meaning is represented in text 
meaning representations (TMRs) that are derived com-
positionally, primarily from meanings of words and 
phrases in the text. Word and phrase meaning is en-
coded in the ontological-semantic lexicon. The underly-
ing ontology is the main metalanguage of lexical mean-
ing specification.
1
 As a result, TMRs largely consist of 
                                                 
1
 Some lexical elements carry grammatical, pragmatic or dis-
course-related meanings that are reflected in TMRs but not in 
terms of ontological concepts. 
instances of ontological concepts. Some of these in-
stances are remembered (as “facts”) and stored in the 
fact repository, FR, a knowledge base of remembered 
ontological instances. Some facts in the fact repository 
are referred to by proper names in texts–personal 
names, toponyms, names of organizations, specific arti-
facts (“the statue of Liberty”), etc. These proper names 
are stored in the onomasticon, the semantic zones of 
whose entries contain a pointer to a corresponding FR 
element.
2
  
The following example illustrates our ontological-
semantic knowledge resources (for a more detailed de-
scription see Nirenburg and Raskin 2003, Chapters 6-
7). The example is much simpler than the sentences 
from real texts with which the ontological-semantic 
analyzer typically works and is used here for pedagogi-
cal reasons and to save space. Consider the input sen-
tence Alex Patrick makes tools. Morphological and syn-
tactic analysis of this input will yield the following 
structure (throughout this paper we use a presentation 
notation that is simplified for readability): 
 
root     make 
 cat   verb 
 tense  present 
 subject    
  root   ‘Alex Patrick’ 
  cat   noun-proper 
 direct-object   
  root    tool 
  cat   noun 
  number  plural 
 
The relevant zones of the ontological-semantic entry for 
the appropriate (first verbal) sense of make are as fol-
lows: 
 
make-v1 
syn-struc 
 root make 
                                                 
2
 State-of-the-art techniques for recognizing named entities 
not in the onomasticon are incorporated in the analyzer; for 
example, lists of personal names from various languages are 
included in the lexicon. Note also the special reference resolu-
tion issues that maintenance of the FR raises.  
Operative Strategies in Ontological Semantics  
 cat  verb 
 subject  
             root $var1  
   cat   noun 
 direct-object  
        root $var2  
  cat   noun 
 sem-struc 
  CREATE-ARTIFACT 
   AGENT ^$var1 
   THEME ^$var2 
 
In the above entry, the variables are used for link-
ing; the caret means ‘the meaning of’; CREATE-
ARTIFACT is an ontological concept (ontological con-
cepts are in SMALL-CAPS); AGENT and THEME are 
among its case roles and are used in this lexicon entry 
to specify selectional restrictions on this sense of make. 
As it happens, the selectional restrictions listed in the 
ontological definition of the concept CREATE-ARTIFACT 
are sufficient for this case of make, so that no further 
specialization or generalization of these restrictions in 
the lexicon is required. 
 
CREATE-ARTIFACT 
       ... 
 AGENT sem HUMAN 
 THEME sem ARTIFACT 
… 
 
In general, constraints in lexicon entries can modify and 
supplant those listed in the ontology; this is a means of 
controlling the proliferation of ontological concepts in 
the system. Lexicon entries for the appropriate senses 
of the other elements of the input sentence are as fol-
lows (we assume that the FR does not include an Alex 
Patrick, so that regular lexicon entries will have to be 
used): 
 
alex-n1 
 syn-struc 
  root alex 
  cat  noun-proper 
 sem-struc 
  HUMAN 
   FIRST-NAME value alex 
   GENDER  sem male  
 
In the above, the filler of the property GENDER is intro-
duced through the sem facet since it is defeasible (the 
lexicon acquirer judged it plausible that there may be 
females called Alex). 
 
patrick-n2  ;this is the last name sense 
 syn-struc 
  root patrick 
  cat  noun-proper 
 sem-struc 
  HUMAN 
   LAST-NAME value patrick  
 
tool-n1 
 syn-struc 
  root tool 
  cat  noun 
 sem-struc 
   TOOL 
 
In the entry for tool-n1, the semantics is simply a uni-
vocal mapping to an ontological concept.  
On the basis of the knowledge from the ontology, 
the lexicon and syntactic analysis, the ontological-
semantic analyzer will produce the following semantic 
dependency that will serve as the basis of the TMR 
(numbers appended to ontological concept names mark 
them as instances of these concepts): 
 
CREATE-ARTIFACT-73 
 AGENT human-209 
 FIRST-NAME  alex 
 LAST-NAME  patrick  
      THEME set-46 
    element-type tool 
    cardinality  > 1 
 
1.1  Disambiguation 
 
A number of important representational and processing 
issues have been omitted in this presentation, notably, a 
discussion of disambiguation heuristics. The first-line 
mechanism of ambiguity resolution in OS is matching 
selectional restrictions. In our example, selectional re-
strictions on the theme of the proposition head matched 
successfully: indeed, tools are artifacts. As to the re-
strictions on the agent, they have been found to be too 
weak to resolve the ambiguity completely: both the last-
name and the first-name (not shown) sense of Patrick fit 
the selectional restrictions on the proposition head (in-
deed, Alex Patrick may be also be a double first name). 
Additional disambiguation means are required in this 
case.  
We have developed two general methods for addi-
tional sense disambiguation: dynamic tightening of se-
lectional restrictions (Mahesh et al., 1997) and deter-
mining weighted distances among ontological concepts 
activated in the input (using the Ontosearch procedure, 
e.g., Onyshkevych, 1997). None of these methods will, 
incidentally, help in our example, so that additional 
heuristic procedures will have to be built for this type of 
ambiguity. Incidentally, such heuristic procedures could 
include evidence from a wide variety of sources, in-
cluding text corpora. Supporting semantic analysis in 
this way should become an important direction of work 
in corpus-oriented computational linguistics (see further 
discussion below).  
Residual ambiguity is one of several possible initial 
outcomes of the analysis process. Figure 1 illustrates 
the entire set of possible outcomes of the analyzer op-
eration. Note that, in principle, the single-candidate 
outcome may prove to be less than ideal–it might sim-
ply reflect errors and omissions in static knowledge re-
sources. Still, at this time deriving a single output 
serves as the halting condition of the analysis process.  
 
 
 
Figure 1. Eventualities in semantic analysis. 
 
If Alex Patrick were the name of a company (cf. 
Merrill Lynch) and were not listed in the FR and there 
were no textual clues (e.g., Inc.) to recognize it as such, 
the system would fail to produce the correct analysis 
(that is, we would end up with zero candidate out-
comes). Currently, the only way to rectify this state of 
affairs is to (manually) add to the fact repository an in-
stance of CORPORATION with the name Alex Patrick and 
all its known property values. However, we are working 
on coercion rules that in the above example would fa-
vor the corporation reading if the event in question had 
the selectional restriction CORPORATION in the appro-
priate case role.  
 
1.2  Multivalued Selectional Restrictions 
 
Returning to our original example, if Alex Patrick is a 
corporation, then the selectional restrictions on CREATE-
ARTIFACT will be violated (a company is not a human). 
As a result, the corporation sense will always lose to the 
human sense because in the latter case selectional re-
strictions do match! We have, naturally, noted that or-
ganizations are often used in texts in positions that are 
“officially” occupied by people; indeed, this is one of 
the most widespread types of metonymy. We therefore 
decided to relax the selectional restrictions in such 
cases by introducing a relaxable-to facet for property 
fillers, in addition to the rigid value and abductively 
overridable sem facets. This amounted to the introduc-
tion of multivalued selectional restrictions. Matches on 
fillers of sem facets are preferred but matches on fillers 
of relaxable-to facets are not discarded as wrong. If, as 
in our example, ambiguity results, additional means of 
its resolution are used.  
 Multivalued selectional restrictions are also used to 
treat metonymy: e.g., the THEME of PLAY-MUSICAL-
INSTRUMENT is constrained to MUSIC-PIECE but the 
standard metonymy of composer name will be also 
noted in the same lexicon entry, together with its ex-
pansion to MUSIC-PIECE with the property AUTHORED-
BY filled by the metonym. 
 Processing unexpected input (including lexically 
unaccounted for metonymies) can be done dynamically 
using Ontosearch. To treat metonymies, Ontosearch 
calculates whether the ontological distance between the 
metonym and the selectional restriction that it must 
match is below a preset threshold (in which case dy-
namic relaxation is allowed and the case is declared 
true metonymy). 
 
1.3  More Expressive Means in OS 
  
The Alex Patrick example illustrated some of the issues 
involved in building a basic semantic dependency. On-
tological semantics, however, goes beyond that limit 
and into what is traditionally covered in pragmatics and 
discourse – on the assumption that, since these types of 
meanings are expressed in language they ought to be 
represented and manipulated. The following example 
illustrates both the treatment of additional types of 
meaning and additional expressive means in OS.  
The English verb prevent has two senses in the on-
tological-semantic lexicon. The first sense is used, for 
example, in The union organizer prevented a strike. 
The basic meaning of prevent-v1 can be roughly 
glossed as “the union organizer carried out an unnamed 
action or actions as a result of which a strike that was 
looming did not materialize.”  
In the sem-struc zone of prevent-v1, the meaning of 
the syntactic subject (^$var1) has selectional restric-
tions on two facets (sem and relaxable-to). The precon-
dition for the preventive event is the potential of the 
thing that is being prevented. The effect is that the thing 
that was prevented did not materialize. This is encoded 
using value ranges of certain modalities of the event. 
Modalities in OS are means of expressing speaker atti-
tudes toward various elements of meaning (or even en-
tire TMRs). Potential modality reflects the probability 
of a certain event or state of affairs taking place; epis-
temic modality measures the degree of factivity of the 
elements in its scope. Modalities are attributed to spe-
cific sources and take their values from abstract value 
ranges between 0 and 1. Epistemic modality 0 means 
that the event did not take place.
3
  
                                                 
3
 Ontological semantics recognizes modalities other than epis-
temic and potential. The inventory of modalities continues to 
grow and develop – witness the differences between the ac-
counts in Nirenburg and Raskin 2003, Section 8.5.3. and in 
Nirenburg et al. (in preparation).  
 
prevent-v1 
    syn-struc 
 root $var0 
    cat  verb 
    subject  
        root  $var1 
        cat  noun 
  direct-object 
  root $var2 
  cat  noun  
    sem-struc 
       EVENT 
     AGENT       ^$var1  
   sem           HUMAN 
   relaxable-to ORGANIZATION 
     
 PRECONDITION    ^$var2 
            sem EVENT 
             modality potential  > .5 
          epistemic < 1 
      
 EFFECT      ^$var2 
    sem EVENT 
            modality epistemic 0   
 
The second sense of prevent is the one used, for exam-
ple, in Negotiations prevented a strike. Here the mean-
ing is expressed as a CHANGE-EVENT: the event that is 
the meaning of the subject (^$var1) caused this change 
event; there was a potential for ^$var2 to take place; as 
a result of this change, ^$var2 has not taken place (its 
epistemic modality is 0). There is no need for overt 
specification of time dependencies, as causality deter-
mines temporal ordering. (Note that the syntactic ana-
lyzer currently used in OS recognizes certain –ing 
forms, e.g., striking, as nouns.) 
 
prevent-v2 
    syn-struc 
 root $var0 
    cat  verb 
    subject  
        root  $var1 
        cat  noun 
  direct-object 
  root $var2 
  cat  noun 
    sem-struc 
        CHANGE-EVENT 
    PRECONDITION    ^$var2 
           sem     EVENT 
                   modality  potential  > .5 
              epistemic < 1 
      
    EFFECT        ^$var2 
    sem   EVENT 
                   modality  epistemic 0   
 
                 CAUSED-BY        
                   ^$var1 
      sem  EVENT 
 
 The above entries are simplified. For instance, we 
did not show the treatment of generic, timeless state-
ments like Being well prepared prevents unpleasant 
surprises, whose meaning specification does not in-
volve creating a specific instance of either of the two 
events but rather a so-called generic ontological in-
stance representing any element of the class (e.g., tigers 
in tigers are ferocious). A special heuristic rule will be 
used in this case to determine whether generic ontologi-
cal instances are appropriate. One clue that casts a vote 
for the generic reading is the present simple form of the 
verb, which is often associated with timelessness. An-
other is the indefiniteness of the direct object. However, 
these clues do not always guarantee that the statement 
is indeed generalized. One way to ascertain this fact is 
to call a special procedure to check whether ^$var2 co-
refers with a specific concept instance either in the 
TMR or in the fact repository.  
 Procedures such as the above are called meaning 
procedures. For example, the meaning of intensifiers 
(e.g., very) is best expressed in terms of a meaning pro-
cedure – take the meaning of the adjective modified by 
very; it will be expressed as a range on some scale, e.g., 
the meaning of dim will be the range (<> 0.1 0.4) on 
the scale of BRIGHTNESS.  Relative values of such prop-
erties are expressed as ranges on abstract scales be-
tween 0 and 1. The expected absolute boundaries for 
such properties will be defined in the corresponding on-
tological concepts. Thus, while the HEIGHT property of 
HUMAN-ADULT may be marked as the range between 
145 and 200 cm (these values are, in fact, overridable), 
the corresponding values for PROFESSIONAL-
BASKETBALL-PLAYER may be between 180 and 230 cm. 
Values on relative scales can, thus, be resolved to 
ranges on absolute scales. The meaning of very roughly 
amounts to narrowing the range toward its extreme. 
 The presence of meaning procedures demonstrates 
that OS combines declarative and procedural semantics:  
on the one hand, knowledge in OS is separated from the 
processing engines; on the other hand, meaning is cal-
culated dynamically.  
 
2.  Some Distinguishing Features of OS 
 
The above, we hope, has presented an informative 
glimpse into the representational, descriptive and proc-
essing concerns of OS. Space restrictions prevent us 
from describing the many remaining elements of and 
decisions taken in OS. At this point, we’ll attempt to 
summarize important preferences and tenets of our “op-
erational philosophy” that distinguish our work from 
other work and generally characterize OS.  
 The overall emphasis in the development of OS is 
given to: 
 
• maintaining and improving a comprehensive mean-
ing extraction and representation system; 
• gradual enhancement of both the breadth and the 
depth of description (which stand in a trade-off re-
lation relative to a given amount of resources) 
through a large-scale descriptive effort; 
• formulation of heuristics for extracting intended 
meanings and supporting task-oriented processing; 
• attaining adequate descriptive coverage of the mass 
of “uninteresting” linguistic phenomena that ac-
count for a huge majority of occurrences in texts. 
 
Relatively lower (though not zero) priorities are ac-
corded to the study of: 
 
• capabilities of the representation medium; 
• computational complexity of the processes; 
• consistency of the knowledge base; 
• individual “interesting” linguistic cases that are dif-
ficult for people to describe and judge (e.g., com-
plex issues in quantification). 
 
Ontological semantics is a comprehensive approach that 
does not relinquish responsibility for attaining its goal 
to other areas of study or rely on prerequisites that are 
either beyond the reach of the current state of the art or 
are expected as a result of large amounts of research by 
others. It is for this reason that OS includes in its pur-
view all the preprocessing stages in text analysis: to-
kenization, morphology and syntax. Moreover, its pur-
view is broader than that of traditional formal or lexical 
semantics: for example, it does not see a need to sepa-
rate pragmatics and discourse into separate disciplines 
(allowing, for example, formal semanticists to bypass 
many kinds of meaning description on account of their 
being outside the purview of their theories).  
 Ontological semantics seeks to incorporate all the 
relevant topics often studied in relative isolation. These 
topics include language- and knowledge-related issues 
such as aspect, modality, time, causality, quantification, 
text-level relations, non-literal meaning, style, deduc-
tive and abductive presuppositions and entailments, 
nominal compounds, prepositional phrase attachment, 
the meaning of adjectives that do not semantically mod-
ify their syntactic governors, etc. They also include ar-
chitecture and control issues connected with optimizing 
the search for the best semantic analysis in the poten-
tially very large candidate space. We strive to develop 
our own accounts (we call them “microtheories”) for 
each such topic exploiting the research findings of oth-
ers and modifying and expanding them in accordance 
with the practical goals and needs of OS. Unfortunately, 
no feasible solutions have been proposed in the litera-
ture for most microtheories that we need: much of the 
reported work is devoted to rather narrow topics (for 
example, meanings of single words at a grain size be-
yond the capabilities – and often the needs – of the ana-
lyzer) and still other proposals invent very complex 
formal notations without any emphasis on heuristics for 
the assignment of values to language phenomena. (In-
deed, one of the knowledge acquisition rules of thumb 
in OS is not to include in the descriptions everything 
that can be said, only those parts that are actually use-
ful.) 
The purpose of all the microtheories is to help de-
rive text meaning. Some microtheories are driven by 
the needs of the meaning representation (its content, not 
format). For example, all TMRs must at least attempt to 
resolve the deictic indices (speaker, hearer, place, time) 
and, more broadly, all references. Other microtheories 
are forward-chaining – they are triggered not by the 
needs of the output but by the occurrences of certain 
lexical and syntactic configurations in the input – for 
example, nominal compounds. Developing microtheo-
ries means acquiring heuristic rules for the treatment of 
specific phenomena. In OS, the strategy is to use any 
and all potentially useful kinds of knowledge in the left-
hand sides of such rules. Some of this knowledge is en-
coded in the static knowledge sources of the system. 
Other sources of heuristics include syntactic or morpho-
logical information in the input and general properties 
of the input document right down to its formatting 
properties. Finally, the heuristic rules can use evidence 
from various co-occurrence measures for subsets of in-
put in a corpus. We believe that supporting heuristic 
processing in a computational-semantic system is one 
of the more useful applications of corpus linguistics.  
 
3.  Balancing Desiderata and Practical Con-
straints 
 
Ontological-semantic descriptive work is guided by the 
desiderata of breadth and depth of coverage. Constrain-
ing these, however, are the twin limitations of human 
resource availability and the computability of knowl-
edge. In this section we present a number of resource 
acquisition choices, each of which is driven by practical 
needs and represents a whole class of phenomena 
whose continued development occupies the day-to-day 
work in OS. We fully understand that it is inappropriate 
to talk about a particular ontology or even a particular 
lexicon as “the correct one.” While some constraints on 
ontology construction may be considered universal, 
there are many other equally acceptable choices in 
specifying the world model to be used both as the meta-
language for text meaning description and as a major 
source for knowledge to support reasoning. 
 Grain size of static knowledge specification. High 
precision in semantic description always incurs in-
creased acquisition time and often increased processing 
challenges as well. For this reason, OS explicitly pre-
fers well selected simplifications, defined as those not 
expected to noticeably impoverish the usefulness of the 
resulting TMR. For example, about in the meaning 
travel about Europe is mapped to the spatial relation 
INSIDE-OF rather than a) forcing the inclusion of a new 
concept with precisely this meaning or b) adding to the 
lexicon entry complex inferences or effects – e.g., see-
ing numerous different places. Similarly, the meaning 
along as in trees grew along the road is mapped to BE-
SIDE, despite the loss of the precondition that the trees 
cover some unspecified stretch of the road interpreted 
by the speaker as significant. 
 Such conscious simplifications – which, we must 
emphasize, are always open to finer representation 
given the needs of an application and the resulting deci-
sion to thus deploy resources – apply to all aspects of 
knowledge acquisition. For example, due to the quag-
mire of semantic and pragmatic issues involved in in-
terpreting conjunctions like and and but when they 
serve to  link clauses, we level their semantics, creating 
a separate TMR for each clause linked by the functional 
binder ‘conjoined’. Contrast this with our treatment of 
since, because, so that, etc., which are less ambiguous, 
more important for reasoning, and are thus rendered by 
the concepts CAUSED-BY, EFFECT, and ENABLEMENT, as 
applicable.  
 Another example of methodical simplification con-
cerns the expression of time. Temporal properties of 
TMR elements are expressed using a very “lean” set of 
parameters: the ontological relations BEFORE (<) and 
LONG-BEFORE (<<) and their inverses AFTER (>) and 
LONG-AFTER (>>), the concepts START-TIME, END-TIME, 
SPEAKER-TIME, DURATION, and time measurement units. 
We also define two meaning procedures – find-anchor-
time and combine-time. find-anchor-time attempts to 
determine the deictic index, or anchor, relative to which 
other times are defined in the text. It uses a variety of 
heuristics, including times of events mentioned in the 
text as well as the dateline of the text, if available. find-
anchor-time may fail to determine the anchor, in which 
case, at present, the time statements in TMRs are made 
relative to an unknown time of speech, t
0
. combine-time 
takes as input an anchor time and an expression denot-
ing a time period, one of whose ends is the anchor, and 
returns the (actual or relative) time of the combination.  
 As an illustration, below are abbreviated meaning 
representations of some time-related word and phrase 
senses:  
 
after-p4 (as in “ten hours after the operation”) 
  time (combine-time ^$var2.time ^$var3 after) 
 
for-p5 (as in “dribbling for thirty seconds”)  
  time (duration $var2 (sem TEMPORAL-UNIT)) 
 
for-p6 (as in “for the time being”)   
  time (find-anchor-time) 
 
Always open to reconsideration, these and other such 
decisions in favor of a coarse grain size, at a minimum, 
help to drive forward the work of supporting our broad-
coverage text processing system.  
 Economy of expressive means. One tenet of OS is to 
avoid a proliferation of ontological concepts, in line 
with the recommendation by Hayes (1979) that the ratio 
of knowledge elements used to describe a set of ele-
ments of the world to the number of these latter ele-
ments must be kept as low as possible. The factors in-
volved in deciding when and when not to introduce a 
new concept constitute something of a cognitive art too 
subtle to capture in a thumbnail sketch of a short article; 
however, a contrastive example might provide some 
insight. 
 Consider spatial after, as in the first house after the 
stop sign. Previous examples showed cases in which an 
ontologically available spatial relation was deemed 
close enough to describe given lexical entities. In this 
case, though, there is no such relation: NEXT-TO, the 
best candidate, is inappropriate since the house could be 
miles after the stop sign. One option would be to de-
scribe the semantics as a script: the first house one sees 
after passing the stop sign during a motion event – but 
this was judged too complex. Instead, we went ahead 
and added a new spatial relation, AFTER-SPATIAL, to ac-
commodate this and synonymous lexical items from all 
the languages that express such a relation (we must em-
phasize that the ontology is language independent). One 
reason we did not expend the effort to more fully de-
scribe the semantics of after is that we have not yet de-
veloped a sophisticated microtheory of spatial relations. 
Therefore, we are not currently poised to make complex 
inferences about spatial relations in texts. However, 
when we do ultimately embark upon a microtheory of 
spatial relations – either stimulated by the demands of 
an application or through the natural overall progres-
sion of ontological-semantic research – we may recon-
sider whether a descriptive rather than an ontological 
representation of after would be most beneficial.  
 The opposite decision was taken in the case of 
about and its synonyms when they convey approxima-
tion. Approximation is too important for reasoning for 
us to make an ontological concept and consider all ap-
proximated entities thus resolved. Instead, we need 
meaning procedures to fix ranges on scales. Preliminary 
analysis suggested that a 7% expansion either way 
around a magnitude works reasonably well in most 
cases: e.g., about 5 gallons is 4.65 to 5.35 gallons; 
about 150 lbs. is 139.5 to 160.5 pounds. However the 
7% rule produces bad results in some instances: e.g., 
resolving about 6 feet tall to 5’7’’-6’5’’ is bad. What 
we need, instead, is 7% of the amount by which peo-
ple’s heights can vary, which is about 2 feet. The 7% 
rule applied to 2 feet yields 5’10.5’’ to 6’1.5’’ – a much 
better reflection of reality. Another failure of the 
straightforward 7% rule involves clock time. For prag-
matic reasons, saying around 10  or around 5:30 im-
plies 10 minutes in either direction, whereas around 
5:15 permits a smaller range and around 7:07 a smaller 
range still. This example of clock time – work on which 
is still under way – shows that some issues considered 
to be of crucial import to TMRs and the reasoning they 
support are immediate priorities in OS despite the over-
head of research and implementation they impose.  
 Including results of abductive reasoning in TMRs. 
To offset any impression that the current state of devel-
opment of OS always opts for shallower analysis, we 
offer an example relating to our evolving treatment of 
verbs that that have a relatively general meaning despite 
the fact that, in a given context, speakers understand 
them to carry more meaning.  
 Consider once again, for example, make in the sense 
of create an artifact (i.e., a man-made physical object). 
This sense of make, which is only one of many, can 
contextually imply different specific types of events 
including but not limited to baking (make a cake), 
sculpting (make a sculpture), recording (make a CD), 
filming (make a film) and knitting (make a sweater). 
The verbal sense of make that covers all of these exam-
ples says that make is a CREATE-ARTIFACT event whose 
subject is HUMAN and whose object is an ARTIFACT.  
 Representing this meaning of make as CREATE-
ARTIFACT captures all the meaning that is explicitly 
conveyed by this lexical item. In that sense, our analy-
sis should be considered complete. However, we as us-
ers of language know much more precisely what activ-
ity is actually carried out in making a cake: by default, 
it’s baking. When we set about to reflect this knowl-
edge in the ontological-semantic knowledge resources, 
we list CAKE as a default THEME of BAKE. Whereas in-
heritance in the ontology allows the match of CAKE 
with CREATE-ARTIFACT, a more exact, constrained 
match obtains between CAKE and BAKE.  (Note that 
CAKE matches CREATE-ARTIFACT because the THEME of 
CREATE-ARTIFACT includes a union of the fillers of 
THEME of all its descendants, which include, among 
many others, BUILD, PREPARE-FOOD, MAKE-LAW, AU-
THOR-EVENT, FILM-EVENT, RECORD-SOUND, RECORD-
TEXT, CREATE-FABRIC-MATERIAL-ARTIFACT, etc., as 
well as their descendants.) 
An important question at this point is whether we 
should act upon the extra knowledge that cakes are usu-
ally baked or simply allow the correct, though more 
vague direct TMR of the input text to stand. The answer 
depends on the circumstances of an application. For ex-
ample, in machine translation, all other things being 
equal, if the direct TMR can be used as the input to 
successful target text generation, there is no need for 
any specialization procedure. There are, however, many 
cases in which such a procedure would improve the 
performance of various analysis tasks. For example, 
such a procedure helps to resolve certain ambiguities, 
e.g., by preferring the oven to the stove sense of range 
in John prepared the cake using the range—see a de-
tailed description in Mahesh et al., 1997.  It also helps 
in the selection and resolution of referring expressions. 
For example, in John made a good cake though the 
oven was not hot enough, the definite referring expres-
sion the oven is legitimately used for a discourse-initial 
reference—and, therefore, no co-reference resolution 
procedure should be called—because the ontology lists 
OVEN as the default instrument of BAKE. This type of 
inference is ultimately made possible by the availability 
of ontological scripts. 
In principle, if we decide to carry out proactive ab-
ductive meaning specialization whenever we come 
across constructions whose verbs have rather general 
meanings (note, incidentally, that many such verbs have 
meanings that are even vaguer and less specific than 
that of make–consider, for example, such verbs as use, 
do, have or get), we can call the meaning procedure 
seek-specification in each such case. However, the ana-
lyzer currently triggers specialization exclusively on an 
as-needed basis, not proactively because proactive evo-
cation of seek-specification will result in conclusions 
that might never be used.  
 Ontological semantics defines many types of mean-
ing procedures, and their application results in both 
more precise and more fine-grain specification of text 
meaning. In fact, when all meaning procedures are ap-
plied, the resulting TMR will contain information that 
was both overtly present in the original text and infor-
mation abductively inferred form the background 
knowledge of the system. We call the former type of 
TMR basic and the latter, extended.  
 
4.  Summary, Status and Future Work 
 
In this paper, we first briefly and informally illustrated, 
using a few annotated examples, the static and dynamic 
knowledge resources of OS. We then presented the 
main motivations and desiderata of our approach and 
discussed issues relating to how to make ontological-
semantic applications feasible through judicious step-
wise enhancement of static and dynamic knowledge 
sources while at all times maintaining a working sys-
tem.  
 The latest implementation of OS uses an ontology 
of about 6,500 concepts each of which is a named col-
lection of property-value pairs (on average, every con-
cept in the current ontology has 16 properties defined 
for it), with the “meta-metalanguage” of properties 
numbering at the time of this writing about 350 rela-
tions and attributes. The current English lexicon is 
about 35,000 entries and growing. There are also Span-
ish and Chinese lexicons, and lexicons for other lan-
guages are under construction. We have also developed 
a collection of ever growing onomasticons. The English 
onomasticon at present contains over half a million en-
tries. In addition to the lexicons and the ontology, the 
static knowledge sources in OS also include morpho-
logical and syntactic grammars to support text analysis 
and generation, as well as a growing fact repository. 
 We are actively developing a variety of ontologi-
cal-semantic microtheories including microtheories of 
time, quantification, approximateness, issues relating to 
sets and ordered lists, several different aspects of refer-
ence, modality, discourse cohesion, script-based abduc-
tive reasoning. We are also developing pedagogical mi-
crotheories devoted to ontology, lexicon and script ac-
quisition. Additional microtheories are planned for de-
velopment as soon as the current ones reach minimum 
utility levels. We continue to augment the basic knowl-
edge resources and are investigating methods of using 
current knowledge resources to speed up acquisition of 
knowledge resources for languages other than those al-
ready in the system.  
 We are also developing a system for the automatic 
extraction of fact repository elements from TMRs. With 
the help of this system, we plan to acquire a large fact 
repository that will be used as the search space in the 
ontological-semantic approach to information extrac-
tion and question answering. 
 Many difficulties still remain in the path toward 
high-quality, broad-coverage extraction and manipula-
tion of meaning from texts. Still, we believe that the 
ontological-semantic approach is on the right path to-
ward this goal – mainly because it is predicated on de-
tailed and flexible descriptive work on language and 
world knowledge, does not rely on unattainable or un-
computable prerequisites and is driven by the require-
ments of specific high-end computational-linguistic ap-
plications.  
  
References 
 
Hayes, P. 1979. The Naive Physics Manifesto. In: Mitchie, D 
(ed.), Expert Systems in the Microelectronic Age. Edin-
burgh: Edinburgh University Press. 
 
Mahesh, K., S. Nirenburg, and S. Beale 1997. If You Have It, 
Flaunt It: Using Full Ontological Knowledge for Word 
Sense Disambiguation. Proceedings of TMI-97, Santa Fe, 
NM, 1-9. 
 
Nirenburg, S., M. McShane and S. Beale (in preparation). Be-
yond Basic Semantic Dependencies I: The Microtheory of 
Modality in Ontological Semantics. 
 
Nirenburg, S. and V. Raskin. 2003. Ontological Semantics. 
MIT Press (forthcoming). 
 
Onyshkevych, B. 1997. An Ontological-Semantic Framework 
for Text Analysis. Unpublished Ph.D. thesis, Center for 
Machine Translation, Carnegie Mellon University, Pitts-
burgh, PA. 
