The Predictive Function of Prenominal Adjectives
Michael Ramscar & Richard Futrell
Stanford University
contact: [email protected]
Introduction
Prenominal adjectives are supposed to add to
or modify the meanings of head nouns, but this
idea is often problematic (Kamp & Partee,
1995; Ramscar et al., 2010). For example, the
most frequent adjectives before puppy are all
redundant:
Using in COCA (Davies, 2009), we examined the probability of
adjectival modification for 20 high-frequency nouns paired with 20
low-frequency nouns of similar semantics.
In Futrell & Ramcsar (2011) we found that frequent nouns
tend to have different gender than their semantic
neighbors. Could some adjectives show the same pattern—
marking the frequent words in a semantic field?
Results
We examined the 300 most frequent adjectives in COCA
and the distribution of nouns after them. For each adjective
A, we found the 200 most frequent following nouns. We
correlated the log frequency of those nouns with their
probability of being modified by the adjective A.
19 of the 20 lower frequency nouns were more likely to be preceded
by an adjective than the corresponding high-frequency noun,
t(19)=4.1312, p<0.001.
Further, we calculated the probability of those nouns being
preceded by the (i.e. not preceded by any adjective).
Relative frequency
Results
Probability of pre-nominal
adjective
For all adjectives examined, there was a negative
correlation between p(following noun) and
p(adjective|following noun). The results for thin are shown
below.
0.6
0.5
0.4
thin (53 / million) 200 most frequent nouns
0.3
0.14
0.2
0.1
probability of being "modified"
0
0.12
1 2 3 4 5 6 7 8 9 1011121314151617181920 1 2 3 4 5 6 7 8 9 1011121314151617181920
* * * * * * * * * * * * * * * * * * * *
Cold appears before the frequent cold drinks, allowing for
efficient discrimination. Further, the nouns it tends to
appear before—water, beer, and drink—are exactly the
nouns marked by das in German (rather than der or die for
most other cold drinks).
thin
0.06
0.04
R2 = 0.6487
0
Figure 3. Relative frequency and the likelihood of being
preceded by a pre-nominal adjective.
0
1
2
3
4
5
6
7
Log frequency and likelihood of pre-nominal ‘modification’ were
negatively correlated, r = -0.624, p<0.0001.
0.6
most frequnt 20 nouns plus matches
0.5
0.5
2
R = 0.3981
0.4
Further work should examine the distribution of adjectives
in a variety of languages and syntactic positions. Searching
for information-theoretic functions has the potential to give a
unified description of disparate phenomena (gender,
adjectives, noun classifiers, order in naming practices).
0.4
A
0.3
R2 = 0.0296
the
0.2
Conclusions
More informative nouns are more likely to appear with
adjectives. This doesn’t make sense if the function of
adjectives is to add detail to meaning. Yet it is perfectly
understandable if the function of adjectives is entropy
reduction in context (Jaeger, 2010).
log frequency
nouns following thin
probability of being "modified"
Figure 2.
Hypothetical
entropy rate with
‘redundant’
prenominal
adjectives.
A
0.08
0.02
0.6
Figure 1.
Hypothetical
entropy rate with
a bare infrequent
noun.
Figure 6. Distribution of cold before cold drink
nouns.
0.1
probability of being preceded by "the"
If adjectives are only used to convey meaning,
then we expect the opposite: since more
infrequent nouns are often more specific, they
need to be preceded by fewer modifying
adjectives.
An Exception – or a System?
PEOPLE
TIME
YEARS
WAY
YEAR
DAY
WORLD
LIFE
MAN
SCHOOL
THINGS
CHILDREN
STATE
WOMEN
FAMILY
STUDENTS
THING
STATES
GOVERNMENT
NIGHT
TOWNSFOLK
DURATION
EPOCHS
HABIT
EPOCH
AFTERNOON
CONTINENT
BEING
CHAP
SEMINARY
DOOHICKEYS
BRATS
MODE
DAMES
KIN
UNDERGRADUATES
DOOHICKEY
MODES
OLIGARCHY
EVENING
We propose that prenominal adjectives are
used to lower the entropy of informative
nouns in context (see Figs. 1 and 2). If this is
the case, then we should find that more
infrequent nouns are more likely to be
preceded by adjectives.
Study 2: Adjectives
Probability of occrence (blue bars) / co-occrence (red bars)
Before puppy:
cute
little
new
…
Study 1: Nouns
0.1
Literature cited
0.3
0
0
0.2
1
2
3
4
5
6
log frequency
0.1
0
0
1
2
3
4
5
6
log frequency
Figure 4. Probability of adjectival modification by log
frequency of nouns.
7
Figure 5. (top) Of the top 200 nouns following
thin, the probability that they will be preceded
by thin, by their frequency. (bottom) The
probability that those nouns will be preceded by
the, by their frequency.
7
Jaeger, T. F. (2010). Redundancy and reduction: Speakers manage syntactic
information density. Cognitive Psychology 61(1), 23-62
Kamp, H. and Partee, B. (1995). Prototype theory and compositionality. Cognition
57, 129-191.
Ramscar, M., Yarlett, D., Dye, M., Denny, K., & Thorpe, K. (2010). Feature-labelorder effects and their implications for symbolic learning. Cognitive Science
34(7), 909–957.
Davies, M. (2009). The 385+ Million Word Corpus of Contemporary American
English (1990-present). International Journal of Corpus Linguistics 14, 159-90.
Acknowledgments
We thank Melody Dye and Dan Jurafsky for help and
advice.