Principled Disambiguation: 
Discriminating Adjective Senses 
with Modified Nouns 
John S. Justeson* 
SUNY Albany 
Slava M. Katz t 
IBM T. J. Watson Research Center 
Recent corpus-based work on word sense disambiguation explores the application of statistical 
pattern recognition procedures to lexical co-occurrence data from very large text databases. In 
this paper we argue for a linguistically principled approach to disambiguation, in which relevant 
contextual clues are narrowly defined, in syntactic and semantic terms, and in which only highly 
reliable clues are exploited. Statistical methods play a definite role in this work, helping to organize 
and analyze data, but the disambiguation method itself does not employ statistical data or decision 
criteria. This approach results in improved understanding of the disambiguation problem both 
in general and on a word-specific basis and leads to broadly applicable and nearly errorless clues 
to word sense. The approach is illustrated by an experiment discriminating among the senses of 
adjectives, which have been relatively neglected in work on sense disambiguation. In particular, 
the paper assesses the potential of nouns for discriminating among the senses of adjectives that 
modify them. This assessment is based on an empirical study of five of the most frequent ambiguous 
adjectives in English: hard, light, old, right, and short. About three-quarters of all instances 
of these adjectives can be disambiguated almost errorlessly by the nouns they modify or by the 
syntactic constructions in which they occur. Such disambiguation requires only simple rules, 
which can be automated easily. Furthermore, a small number of semantic attributes supply a 
compact means of representing the noun clues in a very few rules. Clues other than nouns are 
required when modified nouns are not useable. The sense of an ambiguous modified noun may be 
needed to determine the relevant semantic attribute for disambiguation of a target adjective; and 
other adjectives, verbs, and grammatical constructions all show evidence of high reliability, and 
sometimes of high applicability, when they stand in specific, well-defined syntactic relations to 
the ambiguous adjective. Some of these clues, however, may be hard to automate. 
1. Introduction 
This paper is an application of corpus analysis to an issue in word sense disambigua- 
tion: to what extent adjectives' senses can be separated by the nouns they modify, 
or more generally, by the phrases they modify or take as complements. We are inter- 
ested specifically in principled disambiguation--systematic interpretation using highly 
reliable inferences based on linguistically motivated features. The work is empirical, 
addressing five of the most common, broadly applicable adjectives in English: hard, 
light, old, right, and short. 
Nouns are intrinsically suited for principled disambiguation of adjectives. Most 
adjectives normally designate an attribute of an entity designated by the noun or noun 
* Department of Anthropology, State University of New York, Albany, NY 12222. 
t Current address: Weston Language Research, 138 Weston Road, Weston, CT 06883. 
(~) 1995 Association for Computational Linguistics 
Computational Linguistics Volume 21, Number 1 
phrase they modify. Often, one sense or group of senses designates an attribute that is 
much more typically relevant to the noun's referent. For example, when the adjective 
old characterizes a human being, as in old man or she is old, it usually means 'aged'--not 
'used,' 'of long standing,' or 'former.' In such cases, the mutual relevance of the adjec- 
tive and noun senses is content specific (semantic) rather than word specific (lexical). 
We presume that it is this semantic relation rather than a lexical association that is 
normally involved in disambiguation. 1 One of the issues to be addressed in disam- 
biguation is the passage from word-specific evidence to conceptual representation, a 
problem that we do not pretend to have solved in this paper. 
In addition, we consider the syntactic constructions within which adjectives occur 
and through which the phrases or clauses they modify are determined. For example, 
some adjectives have senses that are not used predicatively; the use of such an adjective 
as a predicate rules out such senses. The syntactic clues we recognize are as reliable 
as the noun clues. 
Noun-based disambiguation of adjectives is of special interest because a selection 
among different attributes (adjective senses) is likely to be sensitive to the attribute 
bundles (noun senses) they characterize. It does turn out that a small number of se- 
mantic features of nouns do provide fairly high coverage and very high reliability 
in adjective sense discrimination. Noun-based disambiguation is a structured form of 
co-occurrence-based disambiguation, various forms of which are prominent in corpus- 
based work. In co-occurrence-based approaches, it is usual to take into account the 
entire set of words in the vicinity of a target (Maarek and Smadja 1989; Yarowsky 
1992). Gale, Church, and Yarowsky (1992) demonstrate that high reliability and cov- 
erage are simultaneously attainable with such an approach. The type of work we are 
pursuing has the potential to be more readily interpretable, though it is more difficult 
to automate. We suspect that the success of the comparison of contexts in bulk is due 
in large part to the effect of a few highly structured types of clues, such as we examine, 
and in part to more diffuse clues of other types. 
We use statistical inference methods as tools for analyzing and attempting to un- 
derstand the problem of disambiguafion and the potential of resources such as modi- 
fied nouns for solving this problem. Methodologically, our statistical analyses and our 
extraction of disambiguating features from the corpus are straightforward, with one 
exception: in order to be able to base inference on already disambiguated subcorpora, 
we had to devise a way to adjust for the bias brought into the sample by the criteria 
through which they had been disambiguated; the required formulas are presented and 
illustrated in the Appendix. 
The structure of the paper is as follows. In Section 2 we describe the notion of a 
word sense indicator; this defines the types of features we consider as clues for disam- 
biguation generally, and how specifically we use nouns as clues to the recognition of 
adjective senses. In Section 3, we describe the type of data we use; by defining adjec- 
tive senses in terms of the meanings of different antonyms, we can take advantage of a 
large database of examples for analysis, examples that are ideally disambiguated by the 
co-occurrence of adjectives and their antonyms in semantically concordant structures. 
Section 4 provides a detailed report on the structure and results of the investigation 
of noun-based disambiguation. Section 5 discusses more complex types of inference 
1 Word-specific relations between adjectives and nouns are idiomatic, non-compositional pairs (so-called 
"freezes") in which the adjective itself has no independent sense, e.g., hard cash and short cut. In some 
cases, such as hard fact, it is difficult to draw the line between a noun-specific sense (here, 
'incontrovertible') and a compositional sense (e.g., 'inflexible, unyielding'); such indeterminacy is of 
course one of the sources of a freeze. 
2 
John S. Justeson and Slava M. Katz Principled Disambiguation 
involved in some noun-based disambiguation and addresses the potential of other 
types of indicators for adjective senses. 
2. Word Sense Indicators 
Our problem is a specific case of the more general problem of finding clues within 
the context of a word that indicate its sense fairly reliably. Content words that have a 
close syntactic relation to one another are useful candidates for examination and are 
intuitively more likely to bear a close semantic relation than words that are near one 
another but are not related syntactically. One much-studied example is the semantic 
relation between a verb and its arguments (e.g., Boguraev et al. 1989; Church and 
Hanks 1989; Braden-Harder 1991; Hindle and Rooth 1991). 
Discrimination among senses of adjectives based on the nouns they modify or of 
which they are predicated has been the subject of less intensive and systematic study. 
Determining the potential of this line of evidence is the focus of this paper. We do this 
by performing a noun-based disambiguation experiment. Certainly, some nouns are 
strongly associated with particular senses of some of the adjectives that modify them. 
This association can be illustrated for the ambiguous adjective old, which has senses 
roughly synonymous with aged, long existing, former, used, and obsolete, using sentences 
from our experimental corpus (see Section 4.1). Two of the nouns most frequently 
modified by old in general texts are man and house. Overwhelmingly, old is used in the 
sense 'aged (not young)" when it modifies man, e.g., in 
The man was very old and very frail, a widower. 
He was a strong old man: he had lived through forty-five years of those wretched 
casseroles, but she missed him already. 
"Guilty!" came the hoarse croaking sounds of the old men. 
In some sentences, in fact, this noun is the only real basis, within the sentence itself, 
for inferring the sense of old: 
The old man answered this time. 
"Leave the other to the old man." 
"All except the old man." 
Man, therefore, can be taken as a fairly good indicator of the 'aged' sense of old. 
Similarly, when old modifies house, it almost always has one of its 'not new' senses, as 
in 
In the fashionable suburb of Kingston, full of beautiful old houses... 
... around the old Holton house he made many improvements... 
He saw the tractors come and tear down the old houses and plow up the land... 
When old modifies house, then, this is a good indication that old is being used in one 
of these senses. So man and house are reliably associated with different senses of old. 
We refer to man and house as indicators for the senses of old. More generally, a 
feature F (such as the modified noun man) that is associated with a target word T 
3 
Computational Linguistics Volume 21, Number 1 
(such as the modifying adjective old) is an indicator for a sense Si (e.g., 'aged') of T if, 
when the feature is present, that sense Si is more likely than the other senses Sj, j -~ i. 
This statement is formalized in the Appendix. 
This characterization of the indicators for word senses also provides flexibility 
regarding sense definition. Not only does it not require a single assessment, once and 
for all, concerning what the senses of a word may be, it requires no commitment to 
the reality of word senses themselves, as classically construed (see Pustejovsky and 
Boguraev 1993). Disambiguation may be pursued relative to many distinct issues, e.g., 
grammatical class, functional role, document topic, or lexical translation equivalent; 
the entities to be discriminated are the effective "senses" being identified. In this 
paper, we disambiguate relative to a pair of word sense groups, operationally by 
disambiguating relative to sense-specific antonyms; an old man, for example, is a man 
who is not young, and an old house is a house that is not new. Such sense distinctions 
have some justification in terms of the semantic organization of adjectives (see Gross, 
Fischer, and Miller 1989). However, our purpose in choosing them was purely for 
convenience in designing an experiment useful for determining the potential of noun- 
based disambiguation of adjectives. 
The characterization of indicators is equally flexible with regard to the domain 
of its own applicability. It need not be assumed that all instances of the target word 
T are to be included in assessing the relative probabilities of different senses. In the 
experiment described here, we are determining the extent to which those senses of an 
adjective that are associated with one antonym can be distinguished from those associ- 
ated with a different antonym according to the nouns that the target adjective modifies. 
We therefore discriminate between just these two sets of senses, which constitute the 
great majority of instances of the targets, and we exclude from both investigation and 
evaluation all instances in which the sense of the target does not fall in one of these 
two groups. We exclude all freezes from consideration as not being legitimate instances 
in which the adjectives actually have a definable sense (see footnote 1). In addition, 
we exclude the minority of instances that have definable senses that do not fall within 
these two groups. For example, short has a sense 'inadequate' that is related histori- 
cally to its dimensional senses; however, this sense does not have a lexically specific 
antonym, whereas the dimensional senses do (long and tall). 
3. Disambiguated Subcorpora 
To extract a reasonable number of nouns that are indicators for the senses of target 
adjectives, one straightforward approach would be to extract a representative sample 
of sentences for each target adjective, to disambiguate each target manually, and to 
extract those nouns that are relatively frequent and that are modified by the target 
in one sense but not in the other. We adopted a different strategy, one that provided 
us with a large set of sentences in which target adjectives could be disambiguated 
automatically and with complete reliability. This strategy involved disambiguation of 
adjectives by their co-occurrence with sense-specific antonyms. 
Antonym co-occurrence is a frequent and pervasive phenomenon, and it takes 
place under highly restricted semantic and syntactic conditions (Charles and Miller 
1989; Justeson and Katz 1991, 1992). An adjective and its antonym refer to opposed 
values of the same attribute. When they modify the same noun in a sentence this is 
the usual case in sentences in which they both occur--this attribute is virtually ensured 
of applying in a consistent way to both instances of the noun. When an adjective like old 
has different senses that are associated with different antonyms (like new and young), 
the adjective in these sentences is disambiguated by its antonym. Thus, in sentences 
John S. Justeson and Slava M. Katz Principled Disambiguation 
in which old and young modify the same noun, e.g., man, old is thereby interpretable 
as 'not young'; in those in which old and new both modify a noun, e.g., house, old is 
thereby interpretable as 'not new.' 
The reason for this effect is easy to observe. Antonyms most often co-occur in 
direct comparisons or in contrastive opposition, directly reflecting both the identity of 
the attribute to which they pertain and the contrast in its value. As illustrated by the 
following sentences from the Brown Corpus (Francis and Kucera 1982), they usually 
occur in otherwise essentially identical phrases: 
Photograph shows the wrong side of work with light strand being picked up under 
dark strand in position to be purled. 
They indicated that no new errors were being made and that all old errors would be 
corrected within 60 days. 
Note how easy it is to find synonyms for the epithet "miser" and 
how hard to find synonyms for "spendthrift." 
When her right hand was incapacitated by the rheumatism, Sadie learned to write 
with her left hand. 
I found myself becoming one of that group of people who, in Carlyle's words, are 
forever gazing into their own navels, anxiously asking am I right, am I wrong? 
Radio broadcasts, however--now that even plain people could afford loud speakers on 
their sets--held old fans to the major-league races and attracted new ones... 
We often say of a person that he looks young for his age or old for his age. 
We refer to this pattern as phrasal substitution. In these cases, the phrases involved 
usually stand in direct semantic opposition (e.g., her right hand .... her left hand). Co- 
occurring antonyms are also frequently joined by and or or, or appear in noun phrases 
joined by prepositions and having the same head noun: 
It was pitiful to see the thin ranks of warriors, old and young... 
That was one more reason she didn't look forward to Cathy's visit, short or long; 
As for this rider, I never saw him before or afterwards and never saw him dismounted, 
so whether he stood tall or short in his shoes, I can't say; 
Skin colors range from white to dark brown, heights from short to tall, hair 
from long and straight to short and tightly curled. 
It was a winter world without details, a world of shapes in an expanse ranging in color 
from light to dark gray. 
But there is no sudden transition from hard rock to soft. 
The chief function of these conjoined and prepositional co-occurrences is to cover the 
range of possible values of an attribute. 
Whether the antonym co-occurrences involve contrastive opposition or a range of 
attribute values, they call forth the semantic dimension designated by the antonym 
pairs and guarantee concordance in adjective sense of the co-occurring antonyms. 
Thus, when an adjective has different sense-specific antonyms, their co-occurrences 
5 
Computational Linguistics Volume 21, Number 1 
as modifiers of different instances of the same noun reliably disambiguate that adjec- 
tive. Five common English adjectives have such antonyms, yielding ten antonymous 
adjective pairs: hard-easy, hard-soft; light-dark, light-heavy; old-new, old-young; right- 
left, right-wrong; and short-long, short-tall. All the example sentences above involve 
one of these ten pairs, and they exemplify the concordance of the antonyms' senses. 
Furthermore, certain departures from the perfect phrasal substitution patterns equally 
constrain the senses of the antonyms to be concordant. These departures include in- 
sertion of adverbial modifiers, substitution of other words besides the antonyms, or 
use of the antonyms to modify noun phrases having the same head noun in phrases 
that otherwise differ from one another. The subcorpora that were used in the work 
reported here consist of those sentences in which a target adjective and its antonym 
modify separate instances of the same noun or clause. 
4. The Experiment 
This section describes our investigation of noun-based disambiguation and its results. 
Subcorpora were extracted for each of the five target adjectives, consisting of sen- 
tences in which the target was disambiguated by its co-occurrence with an antonym 
as modifiers of the same noun (Section 4.1). The tendency toward sense specificity 
of nouns modified by target adjectives in these sentences is demonstrated by show- 
ing that there is little overlap in the set of nouns modified by the target in the two 
antonym co-occurrence subcorpora for that target (Section 4.2). Nouns indicating the 
antonym-specific senses of these targets were then extracted by statistical analysis of 
their sense preferences. It is shown that these indicator nouns are also specific to the 
senses of the target adjectives in the corpus at large by using them successfully in a 
disambiguation procedure applied to 500 randomly selected sentences (Section 4.3). 
Simple, broadly applicable semantic features characterize most of the indicator 
nouns, whereas broadly applicable syntactic features characterize many of their con- 
texts. Together, these features discriminate the target senses, permitting a more com- 
pact and conceptual rather than word-specific representation of the indicators (Sec- 
tion 4.4): about three-quarters of the adjective instances are disambiguated by these 
features, and virtually errorlessly. 
4.1 Acquiring Disambiguated Subcorpora 
Our original study of antonym co-occurrence (Justeson and Katz 1991) was based on a 
version of the Brown Corpus, containing 54,717 sentences; it yields only 57 sentences 
in which both adjectives from any of the ten antonym pairs analyzed in the present 
study were modifying different instances of the same noun. To get enough sentences 
containing antonym co-occurrences of antonyms to address disambiguation issues ad- 
equately, we used the 1.5 million-sentence APHB Corpus. This corpus of 25,000,000 
words was obtained from the American Printing House for the Blind and archived at 
IBM's T.J. Watson Research Center. It consists of stories and articles from books and 
general circulation magazines. 
All sentences containing co-occurrences of the target adjective and each of its 
antonyms were extracted from the APHB Corpus, yielding 4391 sentential co-occur- 
rences. These sentences were manually postprocessed to eliminate all instances in 
which either the target or its antonym was not being used adjectivally. 2 From the 
2 This could have been automated using a parser. Our immediate interest, however, is in discovering actual patterns of usage and not in building an automatic system. We did the work manually to avoid 
John S. Justeson and Slava M. Katz Principled Disambiguation 
remaining sentences, we further extracted a subset of sentences in which both members 
of the pair modify distinct instances of the same noun. This yielded 1487 sentences 
in which at least one of the target adjectives co-occurs with one of its antonyms, with 
both the target and its antonym modifying instances of the same noun that are in 
separate phrases. Some of these sentences have more than one such co-occurrence, so 
these sentences yielded 1535 total co-occurrences. Every one of these co-occurrences 
had sense-concordant antonyms modifying the same noun; any other sense is usually 
semantically incongruous, especially in direct phrasal substitutions. The sentences in 
which co-occurring antonymous adjectives modify the same noun therefore constitute 
a subcorpus in which the ambiguous members of the antonym pairs are discriminated 
relative to their antonym-specific senses. This gave us ten subcorpora, one for each 
antonym pair, of 1535 examples to use as a database for studying the extent to which 
modified nouns disambiguate their modifying adjectives. 
In counting instances of nouns associated with each adjective, elided nouns and 
anaphoric pronouns were resolved (manually) whenever possible, adding to the counts 
for the noun referent, since we are studying the phenomenon of the adjective-noun 
relation. In addition, we stripped morphological suffixes from noun phrases, to recover 
an adjective-noun base. Thus right winger, right fielder, and heavy sleeper are recognized 
as deriving from right wing, right field, and heavy sleep; short-staffed, short-lived, and light 
industrial are recognized as derived from short staff, short life, and light industry. It was 
counted directly only in non-anaphoric usage. The sentences of this subcorpus contain 
1535 such co-occurrences of the target adjectives and their antonyms. (Co-occurrence 
counts for each of the ten antonym pairs are given in Table 3 of the Appendix.) 
4.2 Specificity of Nouns for Adjective Senses in the Disambiguated Subcorpora 
The disambiguated subcorpora can be used to assess the extent to which target adjec- 
tives, when modifying a given noun, are specific to a single sense rather than being 
usable in either sense. They are in fact surprisingly consistent in this regard, as can be 
illustrated for the indicator nouns discussed in Section 2. There it was noted that man 
is an indicator of the 'aged' sense of old (with antonym young) and house for some of 
the 'not new' senses of old (with antonym new). 
The specific association of man with the 'aged' sense of old is reflected in the use 
of this noun in antonymic constructions. The APHB Corpus contains 64 sentences in 
which both young and old modify man; e.g., 
In Bihzad's paintings we see people and animals as individuals--rich men and poor 
men, old and young, the elders in the mosque and the herdsmen camping among their 
horses in the fields. 
Old men saw visions and young men dreamed dreams. 
I am an old man, you a young one;... 
The old man turns to the young one and says: "The time has come for a few 
questions." 
In both the old and the young man this was a breach of habit. 
Most of these sentences involve phrasal substitution patterns typical of antonyms 
generally. In contrast, there is not one sentence in which both old and new modify 
either man or men. 
the systematic errors to which all automatic parsers are subject. 
Computational Linguistics Volume 21, Number 1 
The converse result obtains in the case of house. Four sentences contain instances 
in which old and new both modify house or houses: 
Cast-iron balustrades became the fashion, to be sought out when old houses were 
pulled down and removed to new houses... 
... entire crates of dishes have been smashed when the trailers cross railroad tracks or 
other rough spots located between the old and the new house. 
Fireplaces in the new house, but not in the old one? 
Section 235 (of the Housing Act of 1968) helps families with low and "moderate" 
incomes to buy one- or two-family houses, old or new. 
In contrast, the APHB Corpus contains no sentence in which both old and young modify 
house or houses. 
The specificity of man and house to particular senses of old is typical of nouns 
in these subcorpora. Most nouns by far were modified by the target in only one 
of its senses, in our co-occurrence sentences. We demonstrated this sense specificity 
of modified nouns by compiling all pairs consisting of a target adjective modifying 
the same noun as either of its antonyms. For each of these adjective-noun pairs, we 
determined the number of instances involving each sense; we want to determine the 
extent to which a particular adjective-noun pair tends to occur with only one of the 
two senses. One hundred and eighty-one adjective-noun pairs occur more than once, 3 
with a total of 1096 occurrences. Most by far have n : 0 sense distributions, i.e., the 
pairs occur in only one of the target's senses. Sixty-one adjective-noun pairs, covering 
828 instances, have 4 or more instances each, and thus could admit 2 or more instances 
in their minority sense. Only 4 do have so many minority instances, covering 39 of the 
828 total occurrences. The occurrence of 2 or more minority instances is not mainly a 
frequency effect; these 4 adjective-noun pairings are no more frequent on average than 
those that do not. So adjective-noun pairs do, as a rule, strongly favor one particular 
sense, and this is as true of pairs with many instances as of those with few. 
4.3 Adjective Disambiguation Using Indicator Nouns 
The specificity of nouns in the disambiguated corpus for senses of the target adjectives 
suggests potentially very high reliability for a noun-based procedure to disambiguate 
common adjectives. We evaluate the potential of such a procedure by extracting, from 
the co-occurrence sentences, a set of nouns that are indicators for the senses of the 
target adjectives and applying them to instances of the targets from non-co-occurrence 
sentences in the corpus at large. 
Because adjectives co-occur with their antonyms fairly frequently, it was practi- 
cal to extract disambiguated subcorpora large enough to provide a base for statistical 
inference. However, subcorpora in which most sentences exhibit phrasal substitution 
of antonyms are clearly not representative samples of the use of the target adjectives. 
This raises the possibility that the specificity of nouns for target adjective senses might 
be influenced by the nature of the sentences in which they occur--those that contain 
largely repeated, contrastive structures; we need evidence concerning their sense speci- 
ficity from the corpus at large. This unrepresentativeness also introduces bias into the 
3 If a pair occurs only once, there is no opportunity for its target adjective to appear sometimes in one 
sense and sometimes in another; so such pairs cannot be used in estimating the consistency of sense 
selection by nouns. 
John S. Justeson and Slava M. Katz Principled Disambiguation 
statistical process of inferring sense indicators for the corpus at large from the spe- 
cially selected subcorpora, a bias for which we must correct. The Appendix gives the 
formula needed to project, from the disambiguated subcorpora to the corpus at large, 
the probability of each sense of the target adjective given the noun it modifies. This 
adjustment for bias requires that we know the sense distribution of the target adjec- 
tives in the corpus to which the indicators are intended to apply and the number of 
instances in the subcorpora in which each potential indicator noun is modified by the 
target adjective in each of its senses. 
We therefore extracted from the APHB Corpus a random sample of 100 sentences 
containing adjectival instances of each target adjective, for a total of 500 sentences in 
all. In addition to requiring that the target word functions as an adjective, we exclude 
all freezes, 4 as well as quantificational expressions such as three years old in which the 
target labels an attribute (e.g., age) rather than a value of that attribute (e.g., 'aged'). 5 
The target adjectives in all 500 test sentences were manually disambiguated, both with 
respect to the antonyms and in some cases with respect to other senses not associated 
with either antonym. The distribution of antonym-related senses is given in the third 
column of Table 1. Since we are interested in discriminating between the two antonym- 
related sets of senses of the targets, we limited attention to those instances of that target 
occurring in a sense for which an antonym exists. For this reason, the total number 
of instances is less than 100 for each target adjective, varying from 89 for hard to 99 
for right. After determining the sense distribution for the target in each set, we could 
project which nouns in the subcorpora are likely to be sense indicators for the target 
adjectives in these samples. 
The following nouns were projected to show a preference for one or the other 
sense of the target adjectives that was statistically significant at the .05 level (nouns 
are in roman, classes in italic): 
hard-not easy: it, clauses 
hard-not soft: none 
light-not dark: none 
light-not heavy: cruiser, harness, load 
old-not new: proper names (of things), world, thing, car, way 
old-not young: man, people, proper names (of people), woman, proper names (of 
places), lady, you, wine, person, bull, he, I, one (animate pronoun) 
right-not left: hand 
right-not wrong: he, I, thing, way, what, clauses, answer, proper names (of 
people), act, country, decision, expert, masturbation, note, people, reason, 
technician, that, theory, you 
short-not long: term, syllable, hair, range, run, story 
short-not tall: none 
4 When an adjective in a frozen noun phrase contributes no distinct meaning to the phrase (see 
footnote 1), it is excluded from consideration; such freezes should be found in the dictionary, and their 
identification in text is a separate problem. 5 The total number of sentences containing the target word that were extracted to yield 100 such 
adjectival instances of the target was 160 for hard, 444 for light, 124 for old, 278 for right, and 147 for short. 
Computational Linguistics Volume 21, Number 1 
Table 1 
Coverage and disambiguation error rates for target adjectives in lO0-sentence samples, using 
different indicator sets. 
Syntax and 
Sense Distribution in Indicator Nouns All Nouns Sem. Attributes 
100-Sentence Samples Coverage Errors Coverage Errors Coverage Errors 
Hard Not easy 69 33 0 33 1 66 0 
Not soft 20 0 0 6 0 14 0 
Light Not dark 31 0 0 23 2 17 0 
Not heavy 67 0 0 8 0 19 0 
Old Not new 48 3 0 20 3 38 0 
Not young 44 28 0 36 0 32 2 
Right Not left 22 8 0 20 1 17 0 
Not wrong 77 43 0 47 0 67 0 
Short Not long 86 11 0 38 0 73 0 
Not tall 9 0 0 6 0 5 1 
126 0 237 7 348 3 
Overall 473 
26.6% 0% 50.1% 3.0% 73.6% 0.9% 
The statistical procedure that was used to identify these nouns as indicators is de- 
scribed in detail in the Appendix. The number of significant indicators recovered is 
quite variable, ranging from none for the 'not soft' sense of hard, the 'not dark' sense 
of light, and the 'not tall' sense of short, to 13 for the 'not young' sense of old and 20 
for the 'not wrong' sense of right. 
At this point, we had extracted a small set of statistically significant nouns that are 
projected to be indicators for adjective senses in the random samples. We then iden- 
tified each instance in these samples in which a target adjective modified a projected 
indicator and tested the agreement of the target's sense with that which the noun 
was projected to indicate. This procedure tests the sense specificity of the projected 
indicators in the 100-sentence samples. 
The results of this test appear in Table 1 under the heading Indicator Nouns. The 
indicators do turn out to discriminate as projected between target adjective senses, and 
they do so with 100% reliability. Given this result, the set of indicator nouns can be 
treated as the basis for a disambiguation procedure. The extent of applicability of such 
a procedure can be inferred from the Coverage column, which records the proportion 
of target adjectives that modify projected indicator nouns. Overall coverage is 26.6% 
(see Table 1), rather low for a disambiguation procedure. The amount of co-occurrence 
data available for inference has had a substantial effect on coverage. Five of the ten 
senses are represented by fewer than 100 co-occurrence sentences each, and only one 
of these five yields any coverage at all. The other five senses are represented by more 
than 100 sentences each, and every one provides some coverage of the 100-sentence 
samples. There is therefore every reason to believe that coverage would increase with a 
larger base for inference. While we found a good semantic matching of adjective senses 
with the indicators that were recovered from the co-occurrence sentences, the indica- 
tor selection depended on an arbitrarily selected 5% level of statistical significance. 
We therefore investigated the dependence of performance on the chosen significance 
level. This dependence is highlighted most clearly by comparing the performance of 
10 
John S. Justeson and Slava M. Katz Principled Disambiguation 
the statistically significant indicators, listed above, with that of the nouns from the 
subcorpora that are not significant as indicators of target sense. For this purpose, we 
therefore treated every noun from the co-occurrence sentences as an indicator of the 
sense which that noun is projected to favor in the sample sentences. The results are 
presented under All Nouns in Table 1. Coverage increased from 126 to 237 instances. 
The 111 newly covered instances are incorrectly assigned in only 7 cases; even when ev- 
ery noun from the co-occurrence sentences is treated as an indicator, reliability remains 
high (97.0%). 
The rather high reliability of even those nouns that are not statistically significant 
indicators of adjective sense suggests that in general text as well as in the co-occurrence 
sentences, most nouns are highly specific to the sense of their modifying adjectives. 
For example, not a single color word is a statistically significant indicator for the sense 
of light, although light blue, light brown, light gray, and light green all clearly use light in 
its 'not dark' sense. This example also illustrates that many of the individual nouns 
that we are treating as separate, independent cases actually manifest a smaller number 
of underlying semantic categories, e.g., color. Speakers' knowledge of language must 
somehow encode such cases, with patterns of use of individual nouns in relation to 
these adjectives emerging on the basis of that knowledge. A natural way to pursue 
the necessary revision is in terms of semantic attributes of these nouns, rather than 
in terms of the nouns themselves. We investigate this possibility, introspectively, in 
Section 4.4. This was already done to some extent when proper names were grouped 
together into classes. 
The modified noun is not always relevant to the process of disambiguation, and 
even when the noun is relevant, it is not always sufficient. The observed errors illus- 
trate this (underlined nouns are from the projected indicators): 
"I have something hard to speak," he remarked. 
A spatula is also used for lifting light pieces of food... 
They \[shells (of bullets)\] were small and light, but their turnip shape and radial fins 
made them difficult to conceal... 
The auctioneer.., auctioned off everything, obviously from the estate of an old, dying 
out family, in short order. 
Our spirit is so twisted, torn, because of self, out of its right center, God, and rooted in 
the flesh; the old life is so foul in the sight of God that no patchwork, no mere polishing 
up, no amount of varnish will do. 
The response to such old masters as Michelangelo, Rembrandt and Velasquez was 
and still is instant wonder and delight. 
The two-inch layer of fat that is attached to the inside of the seal's skin is left intact, and 
finally the whole hide is turned right side out. 
What are the sources of these errors? In the first sentence, something is not modified 
by hard at a deep syntactic level; it is instead to speak something that relates directly 
to hard, the surface modified noun being simply irrelevant (see Section 4.4.2). In the 
next case, the noun is not intrinsically irrelevant, but it turns out not to be useful; 
pieces is virtually empty semantically and can be modified by the target adjective 
in either sense (see Section 5.2). The remaining nouns are relevant. Family, life, and 
master are ambiguous, and once the ambiguity is resolved the sense of the modifying 
11 
Computational Linguistics Volume 21, Number 1 
adjective is reliably indicated; this issue is addressed in Section 5.1. Shell and side are 
also relevant to the sense of the adjective, but even when disambiguated themselves, 
further information about the context of light shells and right side is required before the 
sense of the adjective can be resolved. Section 5.2 addresses some of the contextual 
relations between adjectives and noun senses that sometimes resolve adjective sense 
when the intrinsic attributes of the noun sense do not. 
4.4 Generalizing the Indicator Nouns 
The mutual relevance of nouns and adjectives that permits sense disambiguation is 
concept specific rather than word specific. More than 40 nouns that are identifiable as 
indicators of adjective senses reflect a much smaller number of conceptual categories 
that directly relate to these senses. 
4.4.1 Indicator Noun Attributes. The feature human provides a useful and concep- 
tually well-motivated basis for interpreting old, right, and short. With one exception 
(wine), the projected indicator nouns for the 'aged' sense of old--man, people, woman, 
you, he, person, lady, and proper names of people--refer to human beings. Expanding 
to include all nouns from the co-occurrence sentences that substitute young more than 
new for old, almost all added nouns continue to be for human beings, as well as certain 
pronouns (L we, you, he, she and me, us, her, him), to which animals, plants, and body 
parts are added. In the 100-sentence sample, all of the 10 'aged' instances of old that 
were not covered by the indicator nouns refer to members of these categories, 7 of 
them to human beings. Similarly, the correctness sense of right refers both to deci- 
sions and to decision-making entities, the latter primarily human; among 119 different 
nouns modified by right in the co-occurrence sentences, 14 of the nouns modified by 
'not wrong' instances are +human, and all 55 nouns modified by 'not left' instances are 
-human. Finally, in the case of short, the vertical extent feature characterizing its 'not 
tall' sense is appropriate to relatively freestanding entities that are normally vertical, 
and humans are the most talked-about instances of such objects generally. There are 
differences among the different target adjectives in the appropriateness of the feature. 
In the case of old, it is a restricted version of the feature living thing; for right, of 
animate (people and animals); and for short, human beings happen to be a frequent 
instantiation of a verticality feature that is normally appropriate to woody plants and 
to relatively large land animals. 
The feature concrete is also very widely applicable. All indicators of the 'not soft' 
sense of hard are +concrete, so -concrete reliably indicates the 'not easy' sense of 
hard. Because hard in its 'not easy' sense also modifies concrete nouns syntactically, 
on the surface (though not semantically; see Section 5), +concrete does not as reliably 
indicate the 'not soft' sense of hard. Similarly, -concrete indicates the 'not wrong' 
senses of right, the 'not long' senses of short, and the 'gentle' subset of the 'not heavy' 
senses of light. It also indicates 'not new' senses of old, but in this case -concrete is 
simply a special case of -animate. 
Another widely relevant class of indicators are body parts. These indicate the 'not 
young' sense of old and the 'not long' sense of short. A substantial subset of them indi- 
cate the directional sense of right. This sense is associated with horizontally separated 
members of (mostly inherently) paired, repeated entities; it is appropriate to horizon- 
tally separated, paired body parts eyes, ears, thumbs, hands, wrists, arms, legs, feet, 
etc.--which constitute the majority of nouns, by text frequency, that are modified by 
the 'not left' sense of right. For noninherently paired entities, more complex phras- 
ings such as on the right or rightmost are used instead (also on the left or leflmost). Body 
parts are less well represented in the co-occurrence sentences for hard; those that occur 
12 
John S. Justeson and Slava M. Katz Principled Disambiguation 
(shell, palate, tissue) are sense specific for 'not soft,' being specific cases of +concrete 
and subject to the reservations on +concrete indicator nouns noted above. 
Other semantic features are more restricted to individual adjective senses. The 
attribute +color disambiguates about half of the 'not dark' instances of light. Length- 
related ('not long') senses of short are indicated by nouns that are +time period (term, 
period, day, duration, minute, month, night, time, weekend, in the co-occurrence sentences), 
but this attribute is subsumable under -concrete. Text/utterance type (e.g., story, note, 
book, manuscript, monolog, phrase, speech, stanza), though largely subsumable under 
-concrete, often have +concrete realizations (as for book, note, manuscript). 
Some highly specific attributes are clearly relevant; for example, +military entity 
(cruiser, carrier, gun, arms, armor, dragoon, flak, machine gun, missile) indicates the 'not 
heavy' sense of light, and +mental (answer, decision, reason, theory, argument, assump- 
tion, conclusion, conviction, guess, method, notion, opinion, policy, prediction, proposition, 
question) indicates 'not wrong' senses of right. Until many more adjectives have been 
investigated, we avoid introducing overly specific features whose range, applicabil- 
ity, and definition may be unclear. These are perhaps subsumable under more general 
attributes. This is the case for +mental, a special case of a highly reliable and more gen- 
eral indicator, -concrete (see Section 5.2 concerning the more complex case of military 
entity and the relation of text type, which we do use, to the time period attribute). 
In some cases, semantic features such as those discussed above can be used 
straightforwardly as sense indicators, just as nouns were. For example, a +human 
noun indicates the 'not tall' sense of short; a -concrete noun indicates a 'not long' 
sense of short. In other cases, however, the move from specific nouns to semantic fea- 
tures as sense indicators necessitates the formulation of more specific rules for using 
them. For example, -concrete nouns indicate a 'not heavy' sense of light, except that 
-concrete nouns that are +color indicate its 'not dark' sense. This requires a formu- 
lation using either feature combinations or rule ordering. We can implement this by 
first attempting to apply a rule +color ~ 'not dark' and then attempting to apply a 
rule -concrete ~ 'not heavy', to any unresolved cases (Table 2). More complex rules 
can be expected to be required in other cases. 
We do not propose, in general, to extract automatically either semantic general- 
izations like those discussed above, or the rules that use them. Informed introspective 
analysis, aided by a perusal of corpora, seems a surer way toward rule-based for- 
mulations. Human analysis and understanding are simply richer than the mechanical 
statistical tools and data sources presently available for arriving at such rules. 
4.4.2 Syntactic Structure and the Adjective-Noun Relation. When the nouns are rel- 
evant and indicator nouns are readily recovered, Section 4.4.1 shows that coverage can 
be increased by exploiting specific indicator nouns in order to infer or to extract au- 
tomatically general semantic attributes of nouns. For some of the indicators, however, 
generalization properly takes another course, leading not to semantic but to syntac- 
tic cues for sense identification. For example, predicate adjective usage indicates the 
correctness sense of right, which is clearly manifested throughout the APHB Corpus. 
That's not quite right. 
Much thought had gone into that costume, and it seemed just right for a poor man's 
wife. 
We refer to these as cases of a predicative indicator feature. 
The most prominent example in our data is generalization from the non-anaphoric 
it indicator for the 'not easy' sense of hard, which is also applicable for the 'not wrong' 
13 
Computational Linguistics Volume 21, Number 1 
sense of right. This indicator is found in statements of the form It BE ADJ + infinitival 
clause, where BE is a form of the verb to be. Auxiliary verbs, adverbials, or negation 
occurs optionally in these statements, and the infinitival clause need not follow imme- 
diately. It is often deleted, and in such cases an anaphoric pronoun may replace the 
non-anaPhoric it, resolving to a preceding or following clause. Also, the verb BE may 
be deleted and the entire construction subordinated to a higher verb such as seems or 
becomes. We refer to these as cases of an infinitival indicator feature. 
It wasn't hard to find Marietta Price. 
Somehow it has never been hard for me to believe in Francis" wounds... 
It's sometimes hard for a motorist to pass a young fellow standing on the 
edge of a highway. 
Since he's doing this for his physical welfare, it wouldn't be right of me to let 
him be bothered. 
A result of this pattern is that almost any verb will look like an indicator for these 
same senses of hard and right. 6 This appearance, however, is spurious. The proper 
generalization is simply the syntactic construction, including its variants in which 
the non-anaphoric it does not occur. The variants include those in which the entire 
infinitival clause, or a gerundive phrase based on it, serves as the subject of the main 
clause, with hard or right as predicate adjective, 
But becoming more independent is hard for many children. 
and those in which a noun object from the infinitival clause is promoted to serve as 
subject of the verb of being, in place of it. 
Later, Mama may have regretted being married, because Papa was so hard 
to understand. 
Her energy was tremendous, her scruples hard to find. 
Since the adjective is characterizing an action or state of affairs, these cases can be sub- 
sumed under the +activity or -concrete semantic attributes discussed in Section 4.4.1 
as indicators of these senses of hard and right. Even as a premodifier of a noun, the 
adjectives in this construction often relate semantically to the verb phrase, e.g., 
This is a hard program to carry out. 
What is hard is the carrying out of the program, not the program itself. Once again, it 
is the syntactic construction, and not the modified noun, that is the relevant indicator. 
These considerations help us to refine our use of the adjective-noun relation itself 
and to put it on a firmer linguistic footing. The adjective-noun relation is directly per- 
tinent to semantic attributes of both the adjective and the noun only when there is a 
deep syntactic relation between them. In the case of the infinitival and related construc- 
tions, no such relation holds; the noun modified by the adjective at the surface level 
6 A minority of verbs with a specific relation to the opposite sense do exist, e.g., feel relates to the 'not 
soft' sense of hard. Such verbs relate to the alternative sense only when they are outside of the 
infinitival constructions. 
14 
John S. Justeson and Slava M. Katz Principled Disambiguation 
is irrelevant. A principled approach to adjective disambiguation using nouns there- 
fore requires a determination that the adjective modifies the noun at a deep syntactic 
level. It is therefore important to take into account the infinitival construction prior to 
disambiguating any adjective--even those for which it does not constitute an indicator. 
The last two columns of Table 1 present the results of adjective disambiguation by 
a combination of syntactic and semantic indicator attributes. The disambiguating rules 
we used are given in Table 2. The syntactic indicator attributes, predicative and infiniti- 
val, were applied first. Afterward, if a target adjective sense was not resolved, semantic 
indicator attributes were applied; no individual indicator nouns were used. The se- 
mantic attributes that were applied were animate, body part, color, concrete, human, 
and text type; Church and Hanks (1989) had pointed to two of these attributes, person 
and body part (also time, previously mentioned above) in a seemingly casual listing 
of just five attributes potentially useful for describing the lexico-syntactic regularities 
of noun-verb relations. 
Table 1 shows that these few, general attributes cover almost three-quarters of 
all instances of the target adjectives. Disambiguation by these syntactic and semantic 
attributes is effectively as reliable as disambiguation using significant indicator nouns: 
having three apparent errors in disambiguation is not significantly worse than the 
errorless performance of the significant indicator nouns in the 100-sentence samples. 
In fact, under a deeper analysis, these three cases are consistent with the pertinent 
attributes and should not be treated as errors at all. In one sentence, 
In contrast to his rangy sons, he was a short, heavy, oaken-barrel sort of man. 
short modifies sort (-concrete) and was thus assigned the sense 'not long.' However, 
it is actually relevant not to the head of the noun phrase, sort, but rather to man 
(+animate); so treated, short would be correctly assigned to 'not tall.' More complex are 
the two instances of old wine. In the APHB Corpus at large, the new~old contrast applied 
to wine relates chiefly to contexts of production of the wine or of the introduction 
of a type of wine. The young~old contrast relates instead to the maturation of some 
wines, or more generally, to the developmental phases through which wine passes 
while aging over a period of years. It is a cultural (and thus semantic) fact that wines 
and other nonanimate entities that undergo developmental changes and pass through 
maturational stages are treated as living things. Since the two instances of wine in the 
100-sentence samples are of this sort, their old modifiers are properly assigned to 'not 
young'; we assigned them to 'not new' under a literal interpretation of -animate. 
The rules in Table 2 can be easily implemented. The approach presupposes that 
the natural language processing system within which it is applied includes a reliable, 
wide-coverage parser to determine the noun phrase modified by an adjective and 
the head of that noun phrase. The lexical database used by this parser must include 
semantic attribute tags. Most of those used in this paper are already present in some 
available machine-readable dictionaries, such as the Longman Dictionary of Contemporary 
English. Such a disambiguation procedure is capable of disambiguating, with very 
high reliability, about three-quarters of the 100-sentence sample instances of the target 
adjectives we have investigated. 
5. Further Considerations 
Although the sense clues discussed so far can be readily implemented as a disam- 
biguation procedure, about a quarter of all instances of the adjectives under study 
were not covered by the rules presented in Table 2. This section addresses these un- 
15 
Computational Linguistics Volume 21, Number 1 
Table 2 
Disambiguation of adjectives by syntactic and semantic attributes. Rules 
associated with lower numbers are applied before rules associated with 
higher numbers. Rules associated with the same number are unordered 
relative to one another. 
Target Adjective Disambiguating Rules 
Hard 1. +infinitival ~ 'not easy' 
2. -concrete ~ 'not easy' 
+concrete ~ 'not soft' 
1. +color ~ 'not dark' 
2. -concrete ~ 'not heavy' 
1. -animate ~ 'not new' 
+animate ~ 'not young' 
1. +predicative ~ 'not wrong' 
+infinitival ~ 'not wrong' 
2. -concrete ~ 'not wrong' 
+human ~ 'not wrong' 
+body part ~ 'not left' 
1. +human ~ 'not tall' 
-concrete ~ 'not long' 
+body part ~ 'not long' 
+text type ~ 'not long' 
Light 
Old 
Right 
Short 
covered cases. Some are readily characterized in terms of the general approach of the 
previous section; others are more complex. 
The following discussion treats the kinds of properties that systematically relate 
adjective senses to other features of the sentences in which they occur. Unlike the 
previous section, it does not point to any automated procedure to take advantage of 
these properties, or to the role they might play in some more encompassing procedure, 
and it uses coverage and reliability as measures of the actual association of adjective 
senses with other constructs, irrespective of their recoverability from raw text. Thus, 
this section is concerned with the nature of underlying relations--not with formulating 
a disambiguation procedure. 
5.1 Indicator Noun Sense Attributes 
We have found that a substantial proportion of adjectives can be disambiguated by the 
nouns they modify, largely on the basis of general semantic attributes characterizing 
those nouns. These attributes, being semantic, must relate in fact to noun senses and 
not to nouns per se. This issue is finessed, to some extent, in the projected indicator 
nouns and thus in our application of attributes based on them. Some attributes happen 
to apply to all senses of a given noun. For example, in the 100-sentence samples, course 
disambiguates short, though once it is used for 'path' and once for 'class,' because both 
senses are -concrete. Some indicator nouns were extracted, not because the attribute 
applies to all senses of these nouns, but because these nouns are used far more often in 
senses to which an indicator attribute applies than in those to which it does not apply. 
For example, people shows a statistically significant tendency to be associated with the 
'aged' ('not young') senses of old (when people is the plural of person), as judged from 
the co-occurrence sentences, although one instance of the 'of long standing' ('not new') 
senses of old (when people meant 'ethnic group') was also found. 
16 
John S. Justeson and Slava M. Katz Principled Disambiguation 
In many instances in the 100-sentence samples, the noun modified by a target 
adjective was ambiguous with respect to one of the indicator attributes: an indicator 
attribute did characterize some of the noun's common senses, but not others. For 
example, the noun side is projected to occur equally often with each sense of right ('not 
left' 49.9%, 'not wrong' 50.1%). However, this noun has two broad classes of meanings: 
one refers to commitments on issues and is -concrete; the other refers to flanks and 
is +concrete. As expected on the basis of this semantic attribute, right usually means 
'not left' when it modifies side in its (locational) sense 'flank,' but virtually always 
means 'not wrong' when modifying side in its commitment sense. The sense of side is 
therefore a more reliable indicator of the sense of right than is the noun itself. 
If these nouns are disambiguated with respect to the relevant attribute, reliability 
can be increased, as in the case of right side. Coverage will increase as well. Some nouns 
have two or more common senses that disagree in the value of relevant attributes and 
thus were not recovered as indicator nouns; their senses might well be reliable indicator 
features. Disagreements in the value of a semantic attribute for a given noun can even 
be systematic. Any noun N with sense S can be used to mean 'a type of S,' as with 
family doctor in 
Swedes lament the almost total disappearance of the old family doctor. 
When types are construed as -concrete, as when referring to roles, such uses are 
specific to 'not new' senses of old (and to 'not wrong' senses of right). Thus, any noun 
or semantic attribute that is associated with the alternative senses of these adjectives 
would be wrongly interpreted when a 'type of S' usage of the noun is not recognized. 
Another widely pertinent example is the more complex ambiguity in the reference 
of modified nouns for roles or relationships. A role noun almost always refers to an 
individual (+animate) who stands in that relationship to another, as in all the example 
sentences cited below; for example, when the nuns, new and old, filed out of the cloister, 
it was a set of persons and not of relationships who did so. The adjective, however, 
may apply to that individual (a +animate noun sense) or to the role itself (a -animate 
noun sense). Thus, the adjective old may apply either to the relationship or to the role 
designated by doctor, friend, empress, and nun, with old having the sense 'former' or 'of 
long standing': 
... I was with old friends; I had made new friends; and that night I think that I was 
lonelier than ever before. 
It was only to be expected that the lords and ladies of the court would compare the first 
wife and the second, the old empress and the new.., all in favor of the old. 
A prayer, the Bishop's blessing--and the nuns, new and old, filed out to the cloister. 
or to a person having that relationship or role, with old having the sense 'aged': 
•.. he rang to six friends, not too young, not too old, and explained that he'd have to 
postpone their dinner. 
The old doctor and the young doctor rode in silence for two miles and indulged in 
their memories• 
On the contrary, these nuns, young and old, were invariably cheerful and happy, 
almost gay and full of childish fun and laughter... 
17 
Computational Linguistics Volume 21, Number 1 
There is an inherent ambiguity concerning the relation of the adjective old to its noun: 
the referent of the noun is an individual (which is +animate), but it is an associated 
noun sense to which the adjective applies (which may be +animate or -animate) that 
strictly determines the sense of old. In the 100-sentence samples, when old modifies a 
role noun, it always applies in its 'aged' sense to the individual and in its 'former' 
sense to the role. Some role/relationship nouns are used overwhelmingly in their role 
senses (as with friend) or in their personal senses (as with doctor). Otherwise, inferring 
the correct sense for old involves a resolution of the function of the noun. For example, 
an old forest is 'not new' if it has existed for a great period of time and 'not young' if it 
is in an advanced stage of development in the life cycle of forests (cf., the discussion 
of wine in Section 4.4.2). 
This semantic ambiguity in the noun sense to which the adjective applies can 
therefore be resolved by the same rules formulated for unambiguous cases, once the 
relevant noun sense is identified. How to access the relevant noun sense is an unsolved 
problem: the noun's direct referent is an individual, whereas the semantic structure en- 
tailed by the noun is a semantic network, and the adjective may apply to the network's 
noun sense nodes rather than to the noun referent itself. The utility and elegance of 
such semantic representations is suggested by linguistic discussions on lexical seman- 
tics. They have been used with notable success by Fillmore and Atkins (1991), who 
exploit the intricacies of such networks in a now-classic account of the semantics of 
risk, with different nodes of the network providing the locus of what might be dis- 
tilled as the word's distinct senses. Similar network representations are adopted and 
implemented in restricted domains in several computational models (e.g., Sowa 1986). 
The semantics of noun senses therefore relate more specifically and directly to 
adjective senses than do nouns themselves; in fact, 38 (30%) of the 125 cases not 
covered by the rules of Table 2 are resolved when these broader semantic structures 
are used. In most of these cases, noun senses themselves supplied the attributes used 
by the rules of Table 2 to disambiguate adjectives. In other cases, such as role nouns, 
with more complex semantic structures, we are able to resolve the semantic relation 
of adjective and noun, but this ability cannot be captured in rules as simple as those 
of Table 2. 
5.2 Other Indicators of Adjective Senses 
Section 4 showed that indicator nouns and, in particular, certain of their semantic 
features are quite reliable as bases for interpreting the meanings of the adjectives that 
modify them. In some cases, however, nouns provided very little assistance when 
the pertinent semantic and syntactic features do not apply, the same noun is often 
simply consistent with alternative senses. This was systematically true for relationship 
nouns modified by old. Similarly, the adjective light can refer either to weight or to 
color in modifying most concrete nouns. 
In these cases, disambiguation involves words other than the noun that the target 
adjective modifies, standing in other syntactic relations to the target. The effectiveness 
of one such alternative has already been demonstrated--the special case of antonymic 
adjectives. In the special constructions discussed in Section 3 and, in particular, when 
they modify the same noun, they disambiguate one another with almost perfect reli- 
ability. Consider, for example, the sentence 
A piece will seldom bake uniformly, even with the most loving attention--that is, it 
will vary from light to very dark... 
The target adjective light modifies the pronoun it, which refers anaphorically back to 
18 
John S. Justeson and Slava M. Katz Principled Disambiguation 
piece (of food being baked). Light pieces of food may be either 'not dark' or 'not heavy,' 
so the noun provides no substantial aid to interpretation. However, the phrase from 
light to dark secures the intended sense of light. 
Verb senses can relate systematically to adjective senses, because adjectives often 
designate attributes pertinent to the application of the verbal action to/by the referent 
of the modified object/subject noun. Verbs are therefore useful to the interpretation of 
adjectives modifying subject or object nouns. Returning to the difficult case of light, 
one of the 7 sentences in which an error was made when all nouns were treated as 
indicators (see Section 4.3) is 
A spatula is also used for lifting light pieces of food... 
In this sentence, the head noun pieces is irrelevant, essentially empty semantically. The 
critical noun is food, but it is not a directly usable indicator; light food may be either 
'not heavy,' as in this sentence, or 'not dark,' as in the previous example. It is instead 
the verb lifting that provides the best sentence-internal indication of the 'weight' sense 
of light in the example under consideration. The 5 APHB sentences that refer to the 
lifting of a light object all involve the 'not heavy' sense of light. This is not a logical 
requirement--both dark and heavy objects are also said to be lifted. But the weight of 
an object is intrinsically relevant and its color irrelevant to the lifting of ~,hat object; a 
reference to the lifting of a light N in the 'not dark' sense is likely to be misconstrued 
unless additional cues to interpretation are provided. Similar results are found for 
semantically similar verbs. Thus, carry disambiguates light in 
Furniture movers, for example, carry light objects in their hands. 
whereas the modified noun is of no help. Accordingly, physically supports is a seman- 
tic attribute of some indicator verbs for the 'not heavy' sense of light. 
In some cases when a noun or even a noun sense is consistent with more than 
one sense of the target adjective that modifies it, a default target sense may be reliably 
inferred so long as there is no strong counter-evidence in the immediate context. Old 
doctor, for example, means 'aged doctor' in 28 of 30 instances in our corpus. The other 
2 cases are both for old family doctor. In one, the sentence itself makes it clear that 
a generic 'type of doctor' sense was intended (see Section 5.1). Similarly, old means 
'former' in the sentence 
I know that the old family doctor, Dr. Schlomm, always told Manya she could be 
stabilized on medication, that she could be kept under control. 
as shown by the immediately following sentence: 
So did her present doctor. 
A more complex example of default inference is provided by right as a modifier of side. 
The commitment sense of side strongly favors the correctness sense of right, whereas 
the locational senses of side favor the directional senses of right. However, even with 
the locational sense of side, 'not wrong' is not an anomalous usage. Compare, for 
example, the following sentences from the APHB Corpus: 
They were hiding behind the big oak on the left side of the road. 
19 
Computational Linguistics Volume 21, Number 1 
"Can't you see you're on the wrong side of the road?" 
Hoover leaped from his car and ran to the left side of the gangster's car. 
Zaza still stood in the road, on the wrong side of the car. 
In the face of such examples, it becomes difficult to interpret the adjective in such 
sentences as 
But the car, now on the right \[not left\] side of the road, was too late to veer away from 
the second tire... 
He was on the right \[not wrong\] side of the screen, he had an excellent day's work 
behind him, and in two minutes" time he would hear Bing Crosby sing. 
except with respect to the broader discourse context. In our data, locational senses of 
side always involve the directional sense of right in right side of unless there is decisive 
evidence to the contrary in the same sentence or in the immediately surrounding 
discourse. In many of the directional sense uses, we find no overt clue within the 
sentence or the immediate discourse to determine the sense; it appears to be simply 
the assumed interpretation when no specific information contradicts it. In contrast, 
for each of the (6 out of 45) instances in which a correctness sense of right modifies a 
locational sense of side in right side of, there is something explicit in the near context, 
usually in the same sentence. For example, the ambiguous case of the right side of the 
screen, above, is resolved by a preceding sentence (two short sentences intervene): 
Across the centre of the hall hung a screen, and on this screen was being projected a 
motion picture; half the men had to see the picture back to front, because they had to look 
at it from the back \[the wrong side\] of the screen, but nobody minded that very much. 
The directional sense of right cannot be construed as a general default, since the cor- 
rectness sense is far more common overall. Evidently, the locational sense of side of 
is so powerful a semantic indicator for the directional sense of right that people do 
not use right side of in other senses without providing substantial countervaling evi- 
dence for the sense. Relative to a given indicator noun, then, it makes sense to think 
of adjective disambiguation in terms of default interpretations. 
Some noun-based disambiguation of adjectives involves the noun's functionality 
rather than its intrinsic semantic attributes; many such nouns relate to relevant at- 
tributes of indicator verbs. The adjective light, when modifying a +concrete noun, is a 
case in point. Both weight and color characterize all +concrete entities. In fact, how- 
ever, we can reliably determine the sense of light in a number of these cases. Three 
nouns in the co-occurrence sentences emerged as significant indicators for the 'not 
heavy' sense of light--cruiser, load, and harness. It is the weight and not the color of a 
load that is functionally relevant; the thickness and thus weight of a harness that bears 
on the speed and load-bearing potential of the draught animals fitted with it; and the 
heaviness of a vessel that is relevant to the speed, ease of handling, load-bearing po- 
tential, and imperviousness to damage that is pertinent to military cruisers. In fact, if 
it were necessary to specify such entities as being light in color, we would expect that 
their functional specificity for lightness in weight would lead to the use of a more spe- 
cific qualifier, such as light-colored, rather than simply light. Semantic attributes such as 
carried things would characterize nouns such as load, e.g., burden, cargo, or freight, and 
load-bearing equipment or load-relevant equipment would characterize nouns such as 
20 
John S. Justeson and Slava M. Katz Principled Disambiguation 
harness and cruiser. In the sample of sentences containing light, the following might be 
subsumed under such attributes: aircraft, brigade, car, cart, defense, guard, horse, industry, 
package, shell, tank, and weight. Some of these, such as shell and tank, have meanings 
to which the load-bearing issue is not relevant, though in the case of tank the appli- 
cation of the adjective light does appear to restrict its referent to the military vehicle. 
Others, such as car, involve issues of fashion and decoration to which color and thus 
darkness is potentially relevant as well, although the 5 instances of light car(s) in our 
corpus do refer to weight (with dark car(s) referring to a darkened interior). But suit- 
ably constrained either to subsets in which decoration is not a functional value (e.g., 
military~industrial equipment) or by treating feature combinations that include, e.g., 
-decorative relevance, it would be possible if not ideal to handle such nouns in terms 
of attribute values. Thus, some nouns can be disambiguated by relatively narrowly 
defined semantic classes, such as military/industrial equipment, and with high relia- 
bility, leading to something close to lexicalized noun phrases, e.g., light cruiser or light 
industry. But it is in fact the functional relevance of heaviness versus darkness in the 
context of its use that is actually involved. 
A more complex example is provided by those indicators for the 'not long' sense 
of short that are types of texts or utterances--book, manuscript, monologue, note, phrase, 
poem, speech, stanza, story, and syllable. These, like +time period, are largely subsumable 
under the -concrete feature, but there are also +concrete instances of some of these 
nouns, such as book, that still select the 'not long' sense of short. In these instances, it is 
not a physical dimension of the item that is short, and reference to such dimensions in 
the case of book relates instead to the 'not tall' sense of short. Superficially, this might 
support the pertinence of a special attribute, textual, for disambiguating short. How- 
ever, the 'not long' characterization of shortness of texts refers explicitly or implicitly 
to the duration of the performance (e.g., reading or reciting) of the text. Accordingly, it 
is the time period attribute that would appear to be involved in this case---an attribute 
of activities that constitute the typical use of texts, not an attribute of texts themselves. 
The relationship involved in this case is comParable to that discussed by Pustejovsky 
and Boguraev (1993), in which a single sense of fast relates speed of vehicle motion in 
both fast car and fast highway via the qualia structure of lexical entries for highway and 
car. Accordingly, although most instances of nouns for text types can disambiguate 
short by being -concrete, the principled basis for disambiguating the adjective entails 
a more complex type of inference than simple characterization of semantic attributes 
of the modified noun itself. 
6. Concluding Remarks 
For the adjectives and adjective senses under study, it has been demonstrated that the 
noun modified by an adjective provides, in most cases, an extremely reliable indication 
of the sense of that adjective. General semantic attributes of the modified noun provide 
equally reliable and more widely applicable indications of adjective meanings. These 
attributes are conceptually relevant and compact ways of representing the semantic 
relation between adjectives and their modified nouns. When the noun is such that 
it does not disambiguate the adjective, other words in other specific relations often 
provide this information. Once again, these words reflect semantic classes rather than 
specific lexical items and provide a compact and meaningful semantic characterization 
of the relation of the adjective to the related word class. 
This work also supports the view that a small number of close syntactic relations 
channel much of the semantic interpretation involved in disambiguation, at least in 
the case of sense dichotomies in adjectives. We suspect that when this highly struc- 
21 
Computational Linguistics Volume 21, Number 1 
tured type of evidence is removed, the more diffuse kinds of clues implicit in mass 
contextual comparison approaches still have a principled role in disambiguation, e.g., 
by providing a generalized feel for the overall topical preferences of target senses. 
The data on which this study is based were analyzed statistically, and our re- 
sults were summarized quantitatively. These analyses, however, were not provided in 
anticipation of the creation of statistical procedures for automated word sense disam- 
biguation, but rather to help us to explore and to establish a property of language--that 
general semantic attributes of noun senses correlate with the senses of the adjectives 
that modify these nouns. The phenomenon itself can now be formulated in nonquanti- 
tative terms. In addressing or utilizing this phenomenon, corpora remain quite useful, 
supplying examples of word use that can be used for insight and serious discourse 
analysis. 
The main goals of this paper were to investigate conceptual issues in adjective 
disambiguation. The above results, however, may provide a basis for a useful core 
component of an automated procedure that disambiguates adjectives, with the rest 
to be covered by an auxiliary procedure. This core component's coverage of three- 
quarters of adjective instances is lower than can be attained by other approaches, even 
those with quite high reliability. Nonetheless, such a core has practical advantages: 
when it is applicable, it is virtually errorless; and if there are errors, it is known where 
they will be concentrated---in the remaining quarter of cases, those resolved by the 
auxiliary procedure. The full procedure can then attain a 97-98% reliability level even 
if the auxiliary procedure has only 90% reliability. 
The results of this study need to be extended in several ways. First, they should be 
extended to cover a substantial number of adjectives. Any such enterprise always leads 
to improved understanding, but there is also reason to believe that the results will be 
comparable to those we report. The five adjectives investigated in this paper represent 
a fair range in terms of difficulty: the relations between the pairs of senses range from 
barely distinct (short), to interrelated in complex and sometimes subtle ways (old), 
to semantically orthogonal (light). These adjectives are among the most frequent in 
English and, correspondingly, are applicable to a variety of nouns, so they may be 
on average somewhat more useful than most for investigating the interpretation of 
adjectives based upon nouns. It is of course important to investigate types of sense 
discriminations other than the antonymically defined pairs that were imposed by the 
design of this study. 
The results should also be extended to investigate more fully the contributions of 
other types of syntactic relations of adjectives. Although it appears that verb-based 
disambiguation of adjectives can be formulated in the same way as noun-based dis- 
ambiguation, we have not yet systematically studied the contribution of verbs. As in 
the case of nouns, we expect the appropriate representation to be in terms of semantic 
classes rather than sets of words (cf. Levin 1993). It would also be useful to address 
the possible contribution of adjectives other than the sense-specific antonyms. 
Finally, the results need to be extended beyond the class of adjectives. Most directly, 
nouns are probably as reliably disambiguated by adjectives as adjectives are by nouns. 
Appendix A: Extracting Word Sense Indicator Features 
This paper treats the disambiguation of adjectives relative to antonym-specific senses, 
with the nouns they modify as clues suggesting which sense pertains to a given in- 
stance. We model this situation more generally as follows. 
22 
John S. Justeson and Slava M. Katz Principled Disambiguation 
A.1 Definition 
Let T be a target feature of a sentence or other textual unit that is to be disambiguated 
with respect to some set of characteristics $1 ..... Sn; in our study, n is 2, and $1 and 
$2 are two groups of senses, each group relating semantically to a particular antonym 
of a target adjective T. Let F be some feature that can be associated with the target 
feature T in the same textual unit; F may in general be any feature of a sentence or 
other textual unit that contains T, e.g., a semantic attribute, a word, a phrase, or a 
grammatical pattern. In our study, F is a noun modified by a target adjective T. We 
abbreviate the event "sense of T = S" as "S" and "F is in the context of T" as "F." 
Then F provides evidence concerning the sense of T if, for some i and j, the conditional 
probability P(Si I F} is not equal to P{Sj \[ F}. We call F an indicator for the sense Si of 
T if the conditional probability of Si given F is higher than the conditional probability 
of any other sense given F: P{Si \[ F} > P{Sj \[ F} for all j   i. 
A.2 Estimation of P{Si \[ F} 
The desired probabilities P{Si \[ F}, i = 1,..., n, can be estimated from a representative 
sample of sentences with T in a corpus. However, a subcorpus in which T is disam- 
biguated by a particular set of clues may bias the sense distribution; the desired prob- 
abilities differ, in general, from probabilities Q{Si \[ F} defined on such disambiguated 
subcorpora. The formula we derive for P{Si \[ F} takes into account this possible bias. 
In the derivation, we use an intuitively reasonable assumption that the occurrence of 
a potential indicator feature F depends only upon the sense of the target T with which 
it is associated and not upon the fact that it occurs in the subcorpora (in our case, 
primarily a phrasal substitution context of antonym co-occurrence sentences), i.e., that 
Q{F I Si} = P{F I Si}. This need not be true for arbitrary disambiguating features, but 
we treat it as approximately correct in the case of nouns modified by a target adjective. 
Starting with P{Si IF} and applying the identities 
P{Si IF} ---- P{F\[Si}P{Si}p{F} , Q{F\[Si} -- Q{SiQ{Si}\[F}Q{F}, and ~-~P{Sii IF} = 1, 
we obtain for i = 1,..., n 
P{Si} qi , where qi = piQ{Si \[ F} and Pi = Q{Si}" P{Si I r} - Gqk 
k 
The qi are the conditional probabilities Q{Si I F} weighted by the bias ratios Pi defined 
above. When there is no bias, i.e., when Pi = 1 and P{Si} -- Q{Si} for all i, then the 
formula yields simply P{Si l F} = Q{Si l F}. 
For cases with two senses, the formulas are 
P{s, IF} = plQ{S1 \[F} and P{S2 l F} -- p2Q{S2 \[F} plQ{S1 I F} + p2Q{S2 I F} plQ{S1 \] F} q- p2Q{S2 \] F}" 
Most of the target adjectives we investigated have a substantial sense bias, because 
their antonyms co-occur with them at substantially different rates. When all instances 
of F in the disambiguated subcorpora fall in the same sense group, the estimated 
probability for that sense given the feature F is 1, regardless of the values of the 
biases pi; this was the typical case in our experiment because of the specificity of the 
modified nouns (Section 4.2). But whether or not these estimated sense probabilities 
are statistically significant does depend on the Pi values. 
23 
Computational Linguistics Volume 21, Number 1 
We illustrate the use of the formulas for our data. Let T = old, S~ = 'not new,' 
$2 = 'not young,' and F = friend(s). Because we want to estimate the probabilities 
P{Si \[ F} for our sample sentences in particular and not for the corpus at large, it is 
for these sentences that we determine the P{Si} needed for computing the values of 
the pi. The data are given in Table 3. In this table, nl is the number of instances of 
sense $1 of a target in the 100-sentence sample for that target, and n2 is the number 
of such instances for the other sense; P{Si \] F} is the corresponding probability of the 
sense Si of the target in the samples. For our example, 
P { S1 } - nl _ 48 
nl + n2 48 + 44 -- - 0.522 and P{S2} = 1 - P{S1} = 0.478. 
Estimates for Q{S1 } and Q{S2} are obtained from ml and m2, the numbers of instances 
in which the two antonyms modify the same noun as the target in the co-occurrence 
sentences; they are 
ml _ 463 
Q{S1 } - ml q- m2 463 + 266 - 0.635 and Q{S2} = 1 - Q{S2} = 0.365. 
The resulting bias ratios are 
P { S1} 0.522 
Pl -- Q{S1} - 0.635 
P{S2} 0.478 
-- - 0.822 and P2 -- Q{S2} - 0.365 -- - 1.311. 
The bias ratios for all five pairs of adjective senses in our study are given in Table 3. 
The probabilities Q{S1 \[ F} and Q{S2 I F} are estimated from the observed frequencies 
in the co-occurrence sentences. Friend or friends is modified by both old and new in 9 
sentences and by both old and young in 1 sentence, yielding an estimate of 9/10 for 
Q{S1 I F} and 1/10 for Q{S2 \] F}. Substituting these figures into the formula yields the 
estimates 
0.822 x 0.9 
P{S1 I F} = 0.822 x 0.9 + 1.311 x 0.1 = 0.849 and P{S2 I F} = 1 - 0.849 = 0.151. 
According to these probabilities we estimate for the sample sentences, we expect that 
friend should appear to indicate the 'not new' sense of old there, though somewhat 
less than in the co-occurrence sentences themselves. 
A.3 Assessing Statistical Significance of Indicators 
While the estimated probabilities of the senses of old when it modifies friend(s) do 
differ substantially, these estimates are based on only 10 occurrences of friend(s) in 
the co-occurrence sentences. The low number of instances of old modifying friend(s) 
raises the possibility that the difference in estimated probabilities might be spurious, 
a chance deviation. This section explains how we tested this possibility, i.e., how we 
decided whether an observed difference is statistically significant. 
Our null hypothesis is that there is no difference between the probabilities P{S1 I F} 
and P{S2 \[ F}. We test this equiprobability hypothesis using a 5% significance level, 
i.e., we reject it if the probability of so large an observed deviation from this hypothesis 
is 5% or less. Because we did not specify a hypothetical sense preference in advance 
for each word, we use two-tailed tests. 
We frame the null hypothesis of equal sense probabilities P{Si \[ F} in terms of the 
corresponding sense distribution in the sample sentences. We are testing, however, 
24 
John S. Justeson and Slava M. Katz Principled Disambiguation 
Table 3 
Projected probabilities and bias factors for five ambiguous adjectives. S~ and $2 are the 
antonym-related sense groups for the target adjective, in alphabetical order by antonym; the 
mi (i = 1, 2) are the numbers of instances in which the target in sense Si and the corresponding 
antonym co-occur modifying the same noun; the Q{Si} are the relative frequencies, 
mi/(ml + m2); the ni are the numbers of instances of the target adjective in sense Si in the 
100-sentence sample for that target; the P{5i} are the relative frequencies, ni/(nl q- n2); the Pi 
are the bias ratios, P{Si}/Q{Si}; and the Q{Si} are the probabilities projected for the 
co-occurrence sentences under the null hypothesis of equiprobability in the general corpus 
represented by the 100-sentence samples. 
Target 
Adjective ml m2 Q{S1} Q{S2} nl r/2 P{S1} P{S2} pl p2 Q{S1} (7~{$2} 
Hard 14 43 0.246 0.754 69 20 0.775 0.225 3.157 0.298 0.086 0.914 
Light 59 33 0.641 0.359 31 67 0.316 0.684 0.493 1.906 0.794 0.206 
Old 463 266 0.635 0.365 48 44 0.522 0.478 0.822 1.311 0.615 0.385 
Right 312 142 0.687 0.313 22 77 0.222 0.778 0.323 2.487 0.885 0.115 
Short 185 18 0.911 0.089 86 9 0.916 0.084 0.933 1.068 0.518 0.482 
in the co-occurrence sentences. We need to determine the probabilities Q{Si F} for 
the co-occurrence sentences that would correspond to P{Si \] F} = 1/2 in the corpus 
at large. We denote this probability as Q{Si \] F}. We use the same formula for this 
computation as was used in the estimation problem above. This time, however, it is 
the probability in the co-occurrence sentences that we project from an assumed 50-50 
distribution of senses with respect to the noun for the sample sentences; i.e., for cases 
with two senses, we have to solve the simultaneous equations 
1 plQ{S1 \]F} and 1 ~--_ p2(7~{$2 IF} 
2 plQ{St \] F} + p2(~{S 2 \[ F} 2 plQ{S\] \] F} + p2Q{S2 \[ F} 
with Q{S1 \]F} and Q{S2 \[F} as unknowns. The values 
(~{St \] F} - p2 and Q{S2 IF} - pl 
Pl q- P2 Pl -}- P2 
are unique solutions to these equations. 7 
Returning to the example of old friend(s), the null hypothesis is reformulated as 
1.311 
Q{S1 IF} - 0.822 + 1.311 0.615 and Q{S2 IF} = 1 - 0.615 = 0.385. 
We now have a model for the probability of each sense in the co-occurrence sentences, 
on the assumption that sense selection is in fact independent of the modified noun 
7 In the general case, if the target has n senses, the equiprobability hypothesis is expressed by the 
simultaneous equations 
1 pi¢~(s~ I F} 
n = ~,k PkO.{Sk I F} i = 1 ..... n, 
with the unique solutions 
1 
Q(SiJF} - pi i= 1 ..... n. 
k=l ~ 
25 
Computational Linguistics Volume 21, Number 1 
in the corpus at large. In the example of old friend(s), with Q{S1 I F} = 0.615, 6.15 
of the 10 instances of old friend(s) in the co-occurrence sentences are expected to be 
'not new.' The 9 attested instances, then, are somewhat greater than expected under 
the null hypothesis. To decide whether this excess might be a chance deviation from 
expectation, we compute the probability p that 9 or more instances of 'not young' 
would be observed out of 10 instances of old friend(s), if Q{S1 I F} = 0.615. Since we 
are using a two-tailed test with a significance level of 5%, we reject the null hypothesis 
if p < 0.025. 
The formula for computing the one-tailed probability p is a sum of binomial prob- 
abilities. Let St be the sense that we project to be preferred in the corpus at large, given 
that the modified noun is F; i.e., it is the sense $1 if P{S1 I F} > P{S2 I F}, and it is the 
sense $2 otherwise. Then 
P = y~ ml q- m2 k x Q{St IF} k x (1 - Q{St IF}) mlq-m2-k, 
k=mt 
where the mi are the numbers of instances of the target in sense Si that modify F in the 
co-occurrence sentences. In the example of old friend(s), ml +m2 = 10, Q{S1 I F} = 0.615, 
t = 1, and mt = 9, so the probability of observing 9 or more instances of old friend(s) 10 / k0 ) 
in the sense 'not new,' in the co-occurrence sentences, is ~k=9 x 0.615 k x :i 
0.385 l°-k = 0.056 under the null hypothesis. Since p is greater than 0.025, we do not 
reject the null hypothesis; friend(s) is not a statistically significant indicator for either 
sense of old. 
When all instances of F fall in the same sense group (which was typically the 
case in our data), the formula for one-tailed probability p is reduced to the one-term 
expression 
p = Q{St IF} m' 
In the case of old person(s), ml = 0, m2 = 7, P{S1 I F} = O, P{S2 I F} = 1, t = 2, and 
(~{$2 I F} = 0.385; so p = 0.3857 = 0.00125. Since p is less than 0.025, we reject the null 
hypothesis of equiprobability of senses; person(s) is a statistically significant indicator 
for the 'not young' sense of old. 
Acknowledgments 
We thank Branimir Boguraev, Judith 
Klavans, and two anonymous reviewers for 
very helpful comments on earlier versions 
of this paper. 
References 
Boguraev, Branimir; Byrd, Roy; Klavans, 
Judith; and Neff, Mary (1989). "From 
structural analysis of lexical resources to 
semantics in a lexical knowledge base." 
Proceedings, Workshop on Lexical Acquisition. 
International Joint Conference on 
Artificial Intelligence (IJCAI). 
Braden-Harder, Lisa (1991). "Sense 
disambiguatio.n using on-line dictionaries: 
An approach using linguistic structure 
and multiple sources of information." 
Masters thesis, Department of Computer 
Science, New York University, New York. 
Charles, Walter G., and Miller, George A. 
(1989). "Contexts of antonymous 
adjectives." Applied Psycholinguistics 
10(3):357-375. 
Church, Kenneth W., and Hanks, Patrick 
(1989). "Word association norms, mutual 
information, and lexicography." In 
Proceedings, 27th Annual Meeting of the 
Association for Computational Linguistics, 
ACL. 76-83. 
Fillmore, Charles A., and Atkins, Beryl T. 
(1991). "Word meaning: Starting where 
MRDs stop." Plenary paper presented at 
the 29th Annual Meeting of the 
Association for Computational 
Linguistics, Berkeley, California. 
Francis, Winthrop N., and Kucera, Henry 
26 
John S. Justeson and Slava M. Katz Principled Disambiguation 
(1982). Frequency Analysis of English Usage: 
Lexicon and Grammar. Boston: 
Houghton-Miffin. 
Gale, William; Church, Kenneth W.; and 
Yarowsky, David (1992). "Estimating 
upper and lower bounds on the 
performance of word-sense 
disambiguation programs." Proceedings, 
30th Annual Meeting of the Association for 
Computational Linguistics, ACL. 249-256. 
Gross, Derek; Fischer, Ute; and Miller, 
George A. (1989). "Antonymy and the 
representation of adjectival meanings." 
Journal of Memory and Language 
28(1):92-106. 
Hindle, Donald, and Rooth, Mats (1991). 
"Structural ambiguity and lexical 
relations." Proceedings, 29th Annual Meeting 
of the Association for Computational 
Linguistics, ACL. 229-236. 
Justeson, John S., and Katz, Slava M. (1991). 
"Co-occurrences of antonymous 
adjectives and their contexts." 
Computational Linguistics 17(1):1-19. 
Justeson, John S., and Katz, Slava M. (1992). 
"Redefining antonymy: The textual 
structure of a semantic relation." Journal of 
Literary and Linguistic Computing 
7(3):176-184. 
Levin, Beth (1993). English Verb Classes and 
Alternations: A Preliminary Investigation. 
Chicago: University of Chicago Press. 
Maarek, Yoelle, and Smadja, Frank (1989). 
"Full text indexing based on lexical 
relations, an application: Software 
libraries." Proceedings, ACM SIGIR, ACM. 
198-206. 
Pustejovsky, James, and Boguraev, Branimir 
(1993). "Lexical knowledge representation 
and natural language processing." 
Artificial Intelligence. Special Issue on Natural 
Language 63:193-223. 
Sowa, John E (1986). Conceptual Structures: 
Information Processing in Mind and Machine. 
Reading, Massachusetts: Addison-Wesley. 
Yarowsky, David (1992). "Word-sense 
disambiguation using statistical models of 
Roget's categories trained on large 
corpora." In Proceedings, International 
Conference on Computational Linguistics 
(COLING-92). 454-460. 
27 

