THE LOGICAL ANALYSIS OF LEXICAL AMBIGUITY 
Abstract 
Theodes of semantic interpretation which wish to 
capture as many generalizations as possible must 
face up to the manifoldly ambiguous and contextually 
dependent nature of word meaning? In this paper I 
present a two-level scheme of semantic interpretation 
in which the first level deals with the semantic con- 
sequences of'syntactic structure and the second with 
the choice of word meaning. On the first level the 
meanings of ambiguous words, pronominal 
references, nominal compounds and metonomies are 
not treated as fixed, but are instead represented by 
free variables which range over predicates and func- 
tions. The context-dependence of lexical meaning is 
dealt with by the second level, a constraint propaga- 
tion process which attempts to assign values to these 
variables on the basis of the logical coherence of the 
overall result. In so doing it makes use of a set of 
polysemy operators which map between lexical 
senses, thus making a potentially indefinite number of 
related senses available. 
1 INTRODUCTION: LEXICAL 
ASSOCIATION IN A COMPOSITIONAL 
SEMANTICS 
A tenet now held with some force among formal 
semanticists is that the meaning of a complex natural- 
language expression should be a function of just two 
things: the meanings of the parts of the expression 
and the syntactic rule used to form the expression out 
of those parts. Systems such as Montague Grammar 
\[9\] give phrases like "former senator" compositional 
treatments by first translating them to an expression 
of intensional logic, and then giving this expression a 
model-theoretic interpretation in the usual way. The 
practical relevance of this goal to work in natural lan- 
guage processing is clear: for any application domain, 
maximum coverage could be obtained from the same 
domain-independent set of rules, needing only to add 
the relevant entries, with their primitive associations of 
meaning, to the lexicon. 
~The work presented here was supported under OARPA contract 
#N00014-85-C-0016. The views and conclusJons contained in this 
document are those of the author and should not be inte~reted as 
necessarily representing the official policies, either expressed or 
implied, of the Defense Aclvance~ Research Projects Agency or the 
United States Government. 
David Stallard, 
BBN Laboratories Inc. 
10 Moulton St., 
Cambridge, Mass. 
02238 
An obvious technical issue for this program is 
raised by the phenomenon of lexical ambiguity. This 
problem is not one that has been particularly ad- 
dressed in the Montague Grammar literature. The 
most obvious approach is simply to make alternative 
lexical senses separate entries in the lexicon, and to 
allow these disambiguated lexicat items to give rise to 
separate syntactic and semantic analyses. The com- 
putationally unattractive consequences of this are 
quite clear: the same work must be done over again 
for each variant. 
An alternative class of proposals defers the lexical 
part of the analysis until the rest is done. Hobbs 
\[5\] has presented the most detailed general treatment 
of this type to date. This treatment simply associates 
each ambiguous lexical item with the logical disjunc- 
tion of its separate senses. Standard reasoning tech- 
niques (such as theorem proving) can then be ap- 
plied. The problem with this approach is that it is 
simply not correct. This may be most straightfor- 
wardly seen in yes/no questions that contain an am- 
biguity. For example, suppose the the ambiguous 
verb "have" is to be treated as the disjunction of the 
predicates POSSESS, PART-OF, etc. Then the 
answer to the question "Does the butcher have 
kidneys?" must always come out "yes", because the 
second alternative is (assumably) true regardless. 
This method goes wrong because the issue in resolv- 
ing ambiguity is determining which possibility was in- 
tended, not which possibility is true. 
A more correct approach is due to Landsbergen 
and Scha \[8\] and implemented in the PHLIQA1 sys- 
tem. There, the result of semantic interpretation is an 
expression of an ambiguous logical language called 
EFL (for English-oriented Formal Language). During 
semantic interpretation each lexeme is assigned to 
one (possibly ambiguous) descriptive constant of that 
language, which is later mapped, via local translation 
rules, to one or more expressions of an unambiguous 
logical language called WML (for World Model 
Language). The result is a set of complete WML 
translations of the entire EFL expression, from which 
sortally anomalous alternatives are subsequently 
eliminated. 
The PHLIQA1 system, while handling homonymy 
acceptably, does not address the problem of 
polyserny . the presence of an indefinite number of 
related senses for a single word. Consider the 
polysemous lexeme "mouth", which is used differently 
179 
in the phrases "mouth of a person", "mouth of a 
bottle", "mouth of a river", and "mouth of a cave". 
Surely the same logical relationship is not involved in 
each of these cases. Generalizing the meaning of the 
word will not help either, for if we tried to re-define 
"mouth" to mean just any aperture, we would lose our 
ability to refer to human "mouths" independently of 
other parts of the body. Enumerating these separate 
senses with separate translation rules does not look 
like a very promising approach either, since it is not at 
all clear that the list above could not be continued 
indefinitely. The problem with such an approach to 
meaning seems to be that it is too discrete: in linguis- 
tic terms, it does not "capture a generalization". 
This paper presents a method of dealing with lex- 
ical meanings which does seek to capture the 
generalizations implicit in polysemy. The complexes 
of meanings associated with polysemous lexical items 
are generated, structured and extended by a kind of 
"grammar" of word meaning: a set of operators which 
take descriptive constants of a meaning represen- 
tation language onto other descriptive constants or 
expressions of that language. These operations in- 
clude not only metaphorical and metonymic extension 
of the word sense, but "broadening", which allows a 
word to refer to a wider class of items than before; 
"exclusion", which removes from the denotation of a 
word the members of a particular subset thereof; 
"narrowing", which narrows the denotation down to a 
particular subset. Each word is assumed to have a 
core sense (or in the case that it is homonymous, 
several core senses) from which extended senses 
can be derived by recursive application of the 
operators. 
Related to the issue of lexical ambiguity, if tradi- 
tionally studied apart from it. are the problems raised 
by nominal compounds and metonymies. Here the 
problem is determining the binary relation which has 
been "elided" from the utterance. This could in prin- 
ciple be any relation; a translation rule approach can- 
not help here. Novel metaphorical uses of a word, 
such as the substitution of an individual for a whole 
class, will also escape such an approach. The point 
about all three of these phenomena is that they es- 
sentially create new lexical senses. The productive- 
ness of this process suggests that the established 
senses of polysemous lexemes may be generated in 
the same way. 
A key innovation of this work is to treat every non- 
logical word as being potentially ambiguous. Thus 
semantic interpretation initially assigns to each such 
lexical item not an ambiguous constant, but a free 
variable capable of ranging over the appropriate type 
of a higher-order intensional logic \[4\]. These free vari- 
ables are restricted to range not over an explicitly 
enumerated set of logical expressions, but over a 
potentially infinite set of them which is recursively 
enumerable by the polysemy operators. Obviously, 
the core sense itself (and other established senses) 
are not excluded as candidates. A separate con- 
straint propagation stage then assigns appropriate 
descriptive constant values to these variables based 
on the sortal coherence of the whole expression. 
This two-stage method of semantic interpretation 
will be seen to have an advantage over one not dis- 
cussed so far: a single stage method which not does 
not allot a separate role to lexical semantics or pay 
close attention to compositionality, but rather seeks to 
interpret distinct patterns like "mouth of a cave" as a 
whole. Besides suffering from the same lack of 
generality criticised above this latter method en- 
counters difficulty when an ambiguous word-form and 
a pronoun or trace are combined together. A second 
constraint propagation stage enables the dependence 
of word meaning on context - specifically, on the 
meanings of other words and the referents of 
anaphors and deixis in the utterance - to be captured. 
The computational effect is that search can be cut 
down in a space that is essentially a cartesian product 
over the ambiguous elements of an utterance. 
2 THE NOTION OF A "LOGICAL 
VOCABULARY" 
Lexical association cannot be considered apart 
from a notion of "logical" or "conceptual" vocabulary - 
the set of descriptive constants of a logical language 
which are available for making such associations. 
This notion may be identified with the "domain model" 
or "conceptual model" of such systems as PHLIQA1 
\[11\], TEAM \[3\] and IRUS \[1\]. Logical vocabularies, or 
"domains", are what the polysemy operators work 
with. The present section lays down the represen- 
tational structure which the next, dealing with the 
polysemy operators themselves, will make use of. 
Let a "domain" be defined as a set of descriptive 
constants and axioms involving them, subject to three 
conditions: (1) The descriptive constants are such that 
a specification of each of their extensions gives a 
"state of the world" relevant to the domain (2) The 
axioms are such that they constrain which states of 
the world are possible or allowable (3) The axioms do 
not define the constants with the biconditional, but 
with one-way implication only, thus leaving the con- 
stants primitive. If complete definitions of constants 
via lambda-abstraction is allowed it is only as a tech- 
nical convienenca; these are to be regarded as 
"extra". 
The latter condition (3) captures the important fact 
that domains are not definable in terms of other 
domains. Thus expressions cast in logical vocabulary 
O A cannot be directly used to refer to states of affairs, 
etc. expressible only in terms of logical vocabulary 
De. This has an impact for natural language question 
answering systems in which D A is the notions of or- 
dinary language and D B the logical vocabulary of 
some technical domain. In this case, only lexical 
items specially invented for the technical domain 
(such as "JP-5", a particular kind of military jet fuel) 
180 
have an unproblematic lexical association in terms of 
D 8. Obviously not all the words a user employs will 
have this characteristic, nor will all the constants of 
the technical domain be lexicalizable in thisway. In 
other cases the notions of D A will have to be mapped 
to those of D 8, in some way that is not yet specified. 
A common occurrence is for lexical items avail- 
able in regular English to be employed to bridge the 
gap, in such a way as to multiply their effective am- 
biguity. Consider a question seeking to find ships with 
a certain offensive capability: "What ships carry Har- 
poon missiles?". On a literal interpretation of the word 
"carry" the predication of the sentence is satisfied 
whether the ships "carry" the missiles as weaponry or 
as incidental cargo, yet of these only the first alter- 
native is the desired one. If the query were instead 
"What ships carry oranges?" the second alternative is 
the preferable one. The resultant "splitting" of lexical 
senses can be regarded as a form of ambiguity 
generated by the contact between logical domains. 
Other kinds of mapping between notions of dif- 
ferent domains are more complex, not taking place 
along the lines of greater or lesser specificity, but in- 
volving instead another kind of mapping that is really 
tantamount to metaphor. A phrase like "in Marketing", 
for example, is not locative in the literal sense of loca- 
tion in space but rather makes use of a metaphor 
having to do with this notion. Here the initial domain 
is that of space and spatial inclusion, while the final 
one is that of, say, fields of employment or expertise. 
The formal representation of metaphor used in 
this work is that of Indurkhya \[7\]. Indurkhya identifies 
a metaphor with the formal notion of a "T-MAP": a pair 
<F,S> where F is a function mapping descriptive con- 
stants from one domain to another and S is a set of 
sentences which are expected to carry over from the 
first domain to the second. A metaphor is "coherent" 
if the transferred sentences S are logically consistent 
with the axioms of the target domain, "strongly 
coherent" if they already lie in the deductive closure of 
those axioms. 
Depending on the formal language used to 
represent the statements S, one may encounter com- 
putational difficulties (i.e. decidability) with this 
program. One way around this is not to use predicate 
calculus (as Indurkhya does) but a language that is 
more restrictive than predicate calculus. For the price 
of surrending complete expressive power one gains 
the advantage of deductive tractability. 
One system which may be used for this purpose is 
the NIKL \[10\] system, in which only a few types of 
axioms can be encoded. A descriptive constant 
subsumes another of the same complex type if its 
extension is always a superset of the other. Two 
constants are disjoint if their extensions are always 
disjoint. (Note that respective subsumees of the two 
constants "inherit" this disjointness.) Relations of 
more than one argument have sortal (one-place 
predicate) restrictions on their argument, thus stipulat- 
ing that the extension of the relation will always be a 
subset of the cartesian product of the extensions of 
the sorts. Finally, a one-place predicate P restricts a 
binary relation R to be Q if the image under R of each 
member of P's extension is a member of the exten- 
sion of the second one-place predicate Q. In what 
follows I will treat this operation as restricting the form 
that the extension of the relation R may take on, so 
that the placing of constraint P on the first argument 
results in a propagation of the constraint Q on the 
second argument. 
3 THE LEXICAL CONSTRAINT MODULE 
3.1 Overview 
In this section I present a solution to the multiple 
problems of ambiguity posed by a natural utterance. 
Added to the architecture of semantic interpreter, dis- 
course model, lexicon and domain model is a new 
component - the lexical constraint module. It accepts 
from the semantic interpreter a logical form containing 
free variables of higher-order and constructs from it a 
constraint graph structure in which such variables are 
connected in accordance with the syntactic structure 
of the expression. This structure is then used in a 
constraint-propagation process that attempts to as- 
sign descriptive constant values to the expressions. 
The lexicon in this scheme stores for each non-logical 
word an extendable poiysemic complex (or com- 
plexes, in the case of homonymy) of logical associa- 
tions. I shall describe assumptions about the seman- 
tic rule set-up as I go along. 
In making these assignments, the module applies 
a "maxim of coherence". That is, we assume that the 
user will not deliberately speak nonsense to us, use 
terms redundantly, or make use of elaborate means to 
refer to the null set. A coherent outcome is one where 
the descriptive constants being applied to the same 
terms (bound variables and individual constants) are 
not sortaily disjoint. This may not always be achiev- 
able with the core sense of words. When it is not, a 
set of "polysemy operators" is invoked to re-interpret a 
lexical assignment in such a way as to make sense of 
the expression. 
I will first consider an example where no such 
re-interpretation is required. For the utterance "John 
has a car", the following logical form is given as input 
to the constraint module: 
(3 x (=at x) & (have John x)) 
The underlined symbols are the free variables. Sup- 
pose the main verb "have" to be homonymous be- 
tween the various predicates PART-OF, OWN, 
AFFLICTED-WITH. The last of these is eliminable 
because the argument sorts it requires and the the 
sorts given to it do not agree: physical objects and 
b 
181 
diseases are disjoint sets. Such surface inspection of 
argument sorts is not the only source of constraint, 
however. For some relations a particular constraint 
on the first argument causes a constraint on its 
second argument. Thus, the alternative PART-OF is 
eliminable because the parts of an organism must 
themselves be organic material, something clearly 
disjoint with artifacts like cars. The constraint graph is 
now satisfied, and we are left with: 
(3 z (CAR X) & (OWNS JOHN =)) 
3.2 The polysemy operators 
We now proceed to overconstrained cases in 
which potential assignments are in conflict, and re- 
interpretation by the polysemy operators is required. 
For the first pair of such operators, genera/ization and 
exc/usion, we will make use of the Montague Gram- 
mar notion of universal sub/imation \[2\]. A universal 
sublimation of a concept A is just the set of properties 
which are true of all A's members, or: 
(kp (v x A(X) -> P(X})) 
Generalization and exclusion operate upon lexical 
senses by modifying their universal sublimations and 
looking for the alternative meaning (if any) of the word 
that most closely corresponds to this new set. 
As an example of genera/ization, consider the 
phrase "plastic silverware". While in literal terms this 
is oxymoronic, one often sees it used to refer to plas- 
tic eating utensils, and in situations where only these 
items are available, the word "silverware" alone may 
be used to denote them. Obviously for such speakers 
the class EATING-UTENSIL is available as an ex- 
tended and generalized sense of "silverware". The 
initial representation would be: 
(kX (and, (plasti= x) 
(silverwaz. x) ) ) 
A portion of the sublimation of the concept SILVER- 
WARE is the set {MADE-OF-SILVER, EATING- 
UTENSIL}. Of these, it is the first property that is 
disjoint with PLASTIC and a new sublimation is con- 
structed which excludes it. In the partial represen- 
tation above, this new sublimation is just the class 
EATING-UTENSIL itself. 
Exclusion takes a lexical sense onto one from 
which particular sub-senses have been explicitly ex- 
cluded. Consider the sentence "The Thresher is not a 
ship, it's a submarine", or, to be free about its logical 
form: 
(CONTRAST (ship Thresher) 
(submarine Thrlsher) ) 
If we assign the core meanings to these words this is 
nonsensical, since SUBMARINEs are, by definition, 
SHIPs as well. The expression coheres if whatever is 
assigned to ship excludes SUBMARINE. We form a 
partial sublimation {SHIP,~SUBMARINE}, and find 
corresponding to it the alternative sense of "ship", 
SURFACE-SHIP. 
A surprising number of words have such alter- 
native exclusionary senses, among them "axe", where 
HATCHET is excluded; "animal", where HUMAN is 
excluded; and "blue", where TURQUOISE (and other 
off-color shades) is excluded. The phenomenon 
seems to be that a specialized term for some distin- 
guished subset of a concept comes to be the 
preferred term for members of that subset. The all- 
embracing word can still be used, but it comes to 
have a sense which is contrastive with these distin- 
guished subsets. From the impression made by a 
Venn diagram of the set and its excluded subsets we 
might call this "cut-out" polysemy. 
One wonders if certain phenomena which have 
been described as ill-formedness might not in fact be 
instances of this sort of polysemy. Goodman, for in- 
stance, uses the actual word pair "blue" - "turquoise" 
as an example of 
"miscommunication"cite(Goodman85). What seems 
more plausible however, is that the speaker describ- 
ing a turquoise object as "blue" is not really misspeak- 
ing, but is rather using the word "blue" in the more 
inclusive sense which embraces all shades of the 
color. 
Metonymic extension re-interprets a predicate by 
interposing an arbitrary, sortally compatible relation 
between an argument place of the predicate and the 
actual argument. An example can be seen in the 
command "Highlight C3 tracks", where "C3" is a 
predication made of ships and "tracks" are trajectories 
of ship positions, traced out on a screen. Obviously, 
on literal interpretation, this utterance does not 
cohere, since physical objects (ships) and graphical 
objects (tracks) are disjoint. We have: 
(HIGHLIGHT " (kX (=3 X) & (track X) ) ) 
The categories SHIP and TRACK have too many 
clashing properties for generalization or exclusion to 
prevail. Instead, the two clashing elements are recon- 
ciled by finding a function or relation reaching be- 
tween SHIPs and TRACKs (or subsuming categories) 
and metonymically extending one of the items with it. 
The extended meaning of "C3" can be expressed by: 
(kx (3 Y (~a (sHzP Y) 
(SHIP-TRACK Y X) 
(c3 Y) ) ) 
In any usage of the metonomy operation there is a 
choice about which of two clashing elements to ex- 
tend. In this case it would also have been possible to 
have metonymically extended "track" instead of "C3" 
in this example. The resultant expression would then 
be a set of ships instead of tracks - clearly not what is 
wanted here. It would moreover not be an im- 
mediately coherent one itself, since "highlighting" can 
only be done on graphical objects. More importantly, 
it would seem to be that metonomies are less likely to 
182 
shift the head noun meaning, since this changes the 
sortal category of what is being referred to and 
operated upon by the utterance. This seems to be 
particularly strong when the head noun's meaning has 
an underlying functional role, as does "track" in this 
case. 
Note that many words which at first appear to 
have unitary senses are actually better described in 
terms of metonymic complexes. Thus, "window" can 
be used to refer to its constituent pane of glass, its 
sash, or the opening around it. Similar examples can 
be seen in "light", which can be used to refer to the 
actual electromagnetic radiation or the device for 
producing it, and "bank" (in the fiscal sense), which 
can be used to refer to the building or the financial 
institution itself. 
Metaphorical extension operates not by shifting an 
argument place of a predicate, but by shifting the 
predicate itself. Capturing the generality in the mean- 
ing of "mouth" in the example of section 1 involves 
capturing a class of metaphors involving that concept. 
Classes of metaphors are described by the notion of a 
pararneterized T-MAP, in which the mapping function 
F and set of sentences S are not completely specified, 
but may instead have missing elements which must 
be solved for. Let "mouth of the cave" be given by: 
(mouth (iota x (cave x))) 
The functional constant MOUTH is restricted to 
operate on individuals of the class ANIMAL, so the 
above is incoherent on literal interpretation. A 
metaphorical re-interpretation must select certain con- 
stants for the mapping function F and certain facts S 
which carry over to the new domain. Two such facts 
are: 
SUBSUMES (ENCLOSES-SPACE, ANIMAL) 
SUBSUMES (OPENING, MOUTH) 
In this use of the word "mouth" it is operating on in- 
dividuals of the class CAVE instead of ANIMAL. One 
element of the mapping function F is thus the pair 
(ANIMAL,CAVE). In order to determine the relation- 
ship that the word "mouth" really means in the ex- 
ample we must solve for a function variable P which 
MOUTH is mapped to. This function must be sortatly 
coherent with CAVE; it is the righthand member of the 
second ordered pair of the mapping function F. 
The sentences to be transferred are: 
SUBSUMES (ENCLOSES- SPAC~, CAV'E) 
SUBSUMES (OPENING-OF, P ) 
Of these, the first is not only not inconsistent, but true. 
One descriptive constant of the geological domain 
which is obviously not incoherent with CAVE is the 
function CAVE-ENTRANCE. If this function is used in 
place of P the second sentence is satisfied as well. 
An important metric of metaphorical plausibility is 
how much structure in S is transferred from source to 
target domain versus how many descriptive constants 
are mapped via the function F. In the present example 
the ratio is one. Clearly if this ratio is high the 
metaphor is stronger and more plausible; if it is low 
the metaphor is less so. 
3.3 Nominal Compounds 
Nominal compounds are treated by assuming that 
the semantic rules formulate their interpretation with a 
free binary predicate variable standing in for the rela- 
tion which must be determined to complete the inter- 
pretation of the compound. Interpreting the nominal 
compound thus becomes solving for this predicate 
variable. This variable is initially unconstrained ex- 
cept by the sorts of the noun meanings it connects. 
A problem with some nominal compounds is that 
they seem to violate the restrictions imposed by their 
component parts. For example, a "staple gun" is not 
a weapon at all, and would thus on some treatments 
have to be treated either idiomatically or as a com- 
pletely incoherent expression. With the approach 
presented here, however, the polysemy operators can 
be invoked to find a re-interpretation of the words for 
which a solution does exist. The word "gun" can be 
re-interpreted to discard the clashing property of 
shooting bullets only, and to denote in this case the 
wider class of devices that eject objects of whatever 
type. 
An important point about nominal compounds is 
that they cannot be treated extensionally, A soup pot 
is still such whether it currently contains something 
different from soup, or indeed whether it contains any- 
thing at all. Clearly, the relation to be solved for in a 
nominal compound may in general be a non- 
extensional one between kinds, Such a relation may 
in turn have a meaning postulate which dictates which 
actual entities (such as the actual soup) may be re- 
lated at which indices of time. This phenomenon 
would seem to pose a problem for Hobbs and Martin 
\[6\], who view as a sub-problem resolving the refer- 
ence of the "lube oil" in the compound "tube oil alarm". 
One can imagine a "lube oil alarm" which only sounds 
when all the lube oil is gone. 
3.4 Effect on Anaphora Resolution 
Even after syntactic and pragmatic considerations 
have been taken into account, the decision on the 
correct referents for anaphora cannot take place in- 
dependently of considerations of word meaning 
choice. Consider the following two sentences: 
(I) The table i8 in that building 
(2) It im • bank. 
The proper referent of "it" in (2) is constrained by the 
predication made by the ambiguous lexical item 
"bank", namely that it either be a RIVER-BANK or a 
BANK-BUILDING. Neither is sortally coherent with 
TABLE, the referent described by "the table" is elimin- 
183 
able. The only thing left is the individual described by 
"that building" and since BUILDING, being an AR- 
TIFACT, is disjoint with RIVER-BANK, the proper 
sense of "bank" is BANK-BUILDING and the referent 
of "it" is "that building". 
exclusion operator) that an alternative sense of "lion" 
means a male lion only. One should not presume, 
however, that the discovery of new lexical senses will 
occur on a constant basis. The last heuristic above is 
therefore an important one. 
3.5 Algorithm and Heuristics 
The algorithm used by the lexical constraint 
module is a search loop consisting of just three parts - 
tentative assignment, constraint propagation and 
re-interpretation. Qn the first iteration tentative as- 
signment constrains each word-variable with its core 
logical sense, or the set of its core senses if it is 
homonymous. These serve as entry points to the 
polysemy complexes. Variables associated with 
anaphors are initially constrained by whatever prag- 
matic and syntactic (such as C-command) considera- 
tions are seen to apply. The variables associated with 
nominal compounds are initially left unconstrained. 
Thereafter, constraint propagation may end up in 
one of three states: satisfaction, in which case the 
module returns a single logical expression; 
underconstraint, in which case there is an ambiguity 
with which the user must be presented; overconstraint 
in which case re-interpretation is invoked to search for 
an interpretation which coheres. 
The most important issue in performing re- 
interpretation is controlling the process that the sys- 
tem does not "hallucinate" arbitrary meanings into an 
expression. The control heuristics include: 
1. consider overconstrained variables for 
re-interpretation first 
2. prefer generalizations and exclusions 
which modify a small number of 
properties 
3. prefer metaphorical extensions with a 
high ratio of plausibility (as in Sec 3,2) 
and minimize the number of 
"augmentations" and "positings" \['7\] 
4. avoid multiple re-interpretations of the 
same item 
5. prefer re-interpretations to already es- 
tablished polysemous senses instead of 
creating new ones 
In Hobbs' work \[6\] control turns on a notion of a "cost- 
function" associated with the lengths of proofs. The 
notion of "minimality" in that work has some similarity 
to the heuristics above, which seek to avoid arbitrary 
re-interpretations of lexical meanings by prefering 
conservative re-interpretations and discouraging mul- 
tiple ones. 
The creation by the polysemy operators of new 
sense for a word can effectively be regarded as a kind 
of "learning". Thus, given the sentence "That's not a 
lion, that's a lioness" the system could deduce (via the 
4 CONCLUSIONS 
This component will be implemented in a future 
version of BBN's JANUS natural language under- 
standing system. Included in this system will be a 
unification parser with a large grammar and a new 
and improved semantic interpreter. 
I have tried to show how a compositional seman- 
tics need not be incompatible with a context- 
dependent notion of word meaning by making a divi- 
sion of labor between the rule-to-rule translation of 
syntactic structure and the complex semantics of lex- 
ical items. I shall even go so far as to say that such a 
division of labor is neccesary for the compositional 
program to succeed. A component which takes into 
account the creativity of lexical meanings and which 
utilizes knowledge representation and limited in- 
ference not only gives word meaning its proper place 
in a modular system but also has the potential of ex- 
tending coverage and flexibility beyond what is cur- 
rently available in natural language systems. 
Acknowledgements 
I would like to thank Remko Scha for his many 
useful comments on this work. I would also like to 
thank Erhard Hinrichs and Bob Ingria for their com- 
ments and encouragement, and Jessica Handler for 
valuable linguistic data. 
References 
\[1\] Bates, Madeleine and Bobrow, Robert J. 
A Transportable Natural Language Interface 
for Information Retrieval. 
In Proceedings of the 6th Annual Internationa/ 
ACM SIGIR Conference. ACM Special In- 
terest Group on Information Retrieval and 
American Society for Information Science, 
Washington. D.C., June, 1983. 
\[2\] David R. Dowry, Robert E. Wail, and Stanley 
Peters. 
Introduction to Montague Grammar. 
D, Reidel Publishing Company, 1981, 
\[3\] Barbara Grosz, Douglas E, Appelt, Paul Mar- 
tin, and Fernando Pereira. 
TEAM: An Experiment in the Design of Trans- 
portable Natural-Language Interfaces. 
Technical Report 356, SRI International, 
Menlo Park, CA, August, 1985. 
184 
\[4\] 
Is\] 
\[6\] 
(7\] 
C8\] 
\[9\] 
\[10\] 
\[11\] 
Hinrichs, Erhard W. 
A Revised Syntax and Semantics of a Seman- 
tic Interpretation Language. 
1986. 
Hobbs, Jerry R. 
Overview of the TACITUS Project. 
In Proceedings of the DARPA 1986 Strategic 
Computing Natural Language Workshop, 
pages 19-25. The Defense Advanced 
Research Projects Agency, May, 1986. 
Jerry R. Hobbs and Paul Martin. 
Local Pragmatics. 
In Proceedings, UCA/-87. International Joint 
Conferences on Artificial Intelligence, Inc., 
August, 1987. 
To appear. 
lndurkhya, Bipin. 
Constrained Semantic Transference: A Formal 
Theory of Metaphors. 
Technical Report 85i008, Boston University, 
1985. 
Landsbergen, S.P.J. and Scha, R.J.H. 
Formal Languages for Semantic Represen- 
tation. 
In Allen and Petofi (editors), Aspects of 
Automatized Text Processing: Papers in 
Text/inguistics. Hamburg:Buske, 1979. 
Montague, R. 
The Proper Treatment of Quantification in Or- 
dinary English. 
In Jo Hintakka, J.Moravcsik and P.Suppes 
(editors), Approaches to Natural Lan- 
guage. Proceedings of the 1970 Stanford 
Workahip on Grammar and Semantics, 
pages 221-242. Dordrecht: D.Reidel, 
1973. 
Moser, Margaret. 
An Overview of N/KL. 
Technical Report Section of BBN Report No. 
5421, Bolt Beranek and Newman Inc., 
1983. 
Scha, Remko J.H. 
Logical Foundations for Question-Answering. 
Phillips Research Laboratories, Eindhoven, 
The Netherlands, 1983. 
185 
