Detecting the Countability of English Compound Nouns Using 
Web-based Models 
Jing Peng 
Language Media Laboratory  
Hokkaido University 
Kita-14, Nishi-9, Kita-ku,  
Sapporo, JAPAN  
pj@media.eng.hokudai.ac.jp 
Kenji Araki 
Language Media Laboratory  
Hokkaido University 
Kita-14, Nishi-9, Kita-ku,  
Sapporo, JAPAN  
araki@media.eng.hokudai.ac.jp
 
Abstract 
In this paper, we proposed an approach 
for detecting the countability of English 
compound nouns treating the web as 
a large corpus of words. We classified 
compound nouns into three classes: 
countable, uncountable, plural only. 
Our detecting algorithm is based on 
simple, viable n-gram models, whose 
parameters can be obtained using the 
WWW search engine Google. The de-
tecting thresholds are optimized on the 
small training set. Finally we 
experimentally showed that our 
algorithm based on these simple 
models could perform the promising 
results with a precision of 89.2% on the 
total test set. 
1 Introduction 
In English, a noun can be countable or uncount-
able. Countable nouns can be "counted", they 
have a singular and plural form. For example: an 
apple, two apples, three apples. Uncountable 
nouns cannot be counted. This means they have 
only a singular form, such as water, rice, wine. 
Countability is the semantic property that de-
termines whether a noun can occur in singular 
and plural forms. We can obtain the information 
about countability of individual nouns easily 
from grammar books or dictionaries. Several 
researchers have explored automatically learn-
ing the countability of English nouns (Bond and 
Vatikiotis-Bateson, 2002; Schwartz, 2002; 
Baldwin and Bond, 2003). However, all the pro-
posed approaches focused on learning the 
countability of individual nouns. 
A compound noun is a noun that is made up 
of two or more words. Most compound nouns in 
English are formed by nouns modified by other 
nouns or adjectives. In this paper, we concen-
trate solely on compound nouns made up of only 
two words, as they account for the vast majority 
of compound nouns. There are three forms of 
compound words: the closed form, in which the 
words are melded together, such as “songwriter”, 
“softball”, “scoreboard”; the hyphenated form, 
such as “daughter-in-law”, “master-at-arms”; 
and the open form, such as “post office”, “real 
estate”, “middle class”.  
Compound words create special problems 
when we need to know their countability. Ac-
cording to “Guide to English Grammar and 
Writing”, the base element within the compound 
noun will generally function as a regular noun 
for the countability, such as  “Bedrooms”. How-
ever this rule is highly irregular. Some uncount-
able nouns occur in their plural forms within 
compound nouns, such as “mineral waters” (wa-
ter is usually considered as uncountable noun). 
The countability of some words changes when 
occur in different compound nouns. “Rag” is 
countable noun, while “kentish rag” is uncount-
able; “glad rags” is plural only. “Wages” is plu-
ral only, but “absolute wage” and “standard 
wage” are countable. So it is obvious that de-
termining countability of a compound noun 
should take all its elements into account, not 
consider solely on the base word.  
The number of compound nouns is so large 
that it is impossible to collect all of them in one 
103
dictionary, which also need to be updated fre-
quently, for newcoined words are being created 
continuously, and most of them are compound 
nouns, such as “leisure sickness”, “Green fam-
ine”.  
Knowledge of countability of compound 
nouns is very important in English text genera-
tion. The research is motivated by our project: 
post-edit translation candidates in machine 
translation. In Baldwin and Bond (2003), they 
also mentioned that many languages, such as 
Chinese and Japanese, do not mark countability, 
so how to determine the appropriate form of 
translation candidates is depend on the knowl-
edge of countability. For example, the correct 
translation for “发育性痛
1
” is “growing pains”, 
not “growing pain”.  
In this paper, we learn the countability of 
English compound nouns using WWW as a 
large corpus. For many compound nouns, espe-
cially the relatively new words, such as genetic 
pollution, have not yet reached any dictionaries. 
we believe that using the web-scale data can be 
a viable alternative to avoid the sparseness prob-
lem from smaller corpora. We classified com-
pound nouns into three classes: countable (eg., 
bedroom), uncountable (eg,. cash money), plural 
only (eg,. crocodile tears). To detect which class 
a compound noun is, we proposed some simple, 
viable n-gram models, such as freq(N) (the fre-
quency of the singular form of the noun) whose 
parameters’ values (web hits of literal queries) 
can be obtained with the help of WWW search 
engine Google. The detecting thresholds (a noun 
whose value of parameter is above the threshold 
is considered as plural only) are estimated on the 
small countability-tagged training set. Finally 
we evaluated our detecting approach on a test 
set and showed that our algorithm based on the 
simple models performed the promising results.  
Querying in WWW adds noise to the data, 
we certainly lose some precision compared to 
supervised statistical models, but we assume that 
the size of the WWW will compensate the rough 
queries. Keller and Lapata (2003) also showed 
the evidence of the reliability of the web counts 
for natural language processing. In (Lapata and 
Keller, 2005), they also investigated the count-
ability leaning task for nouns. However they 
                                                           
1
 “发育性痛”(fa yu xing tong) which is Chinese compound 
noun means “growing pains”. 
only distinguish between countable and un-
countable for individual nouns. The best model 
is the determiner-noun model, which achieves 
88.62% on countable and 91.53% on uncount-
able nouns.  
 In section 2 of the paper, we describe The 
main approach used in the paper. The prepara-
tion of the training and test data is introduced in 
section 3. The details of the experiments and 
results are presented in section 4. Finally, in sec-
tion 5 we list our conclusions. 
2  Our approach 
We classified compound nouns into three 
classes, countable, uncountable and plural only. 
In Baldwin and Bond (2003), they classified 
individual nouns into four possible classes. Be-
sides the classes mentioned above, they also 
considered bipartite nouns. These words can 
only be plural when they head a noun phrase 
(trousers), but singular when used as a modifier 
(trouser leg). We did not take this class into ac-
count in the paper, for the bipartite words is 
very few in compound nouns. 
 
C-noun
 
 
 
 
 
 
 
 
 
 
 
 
Figure 1. Detecting processing flow 
For plural only compound noun, we assume 
that the frequency of the word occurrence in the 
plural form is much larger than that in the singu-
lar form, while for the uncountable noun, the 
frequency in the singular form is much larger 
than that in the plural form. The main processing 
flow is shown in Figure 1. In the figure, “C-
noun” and “Ns” mean compound noun and the 
plural form of the word respectively. 
“F(Ns)>>F(N)” means that the frequency of the 
plural form of the noun is much larger than that 
of the singular form. 
F(N)>>F(Ns)
plural only
uncountablity
countablity
F(Ns)>>F(N)
Y
N
Y
N
104
Our approach for detecting countability is 
based on some simple unsupervised models.  
θ≥
)(
)(
Nf
Nsf
             (1) 
In (1), we use the frequency of a word in the 
plural form against that in the singular form. θ  
is the detecting threshold above which the word 
can be considered as a plural only. 
θ≥
),(
),(
Nsmanyf
Nmuchf
  (2) 
In (2), we use the frequency of a word in the 
singular form co-occurring with the determiner 
“much” against the frequency of the word in the 
plural form with many, if above θ , the word can 
be considered as uncountable word. (2) is used 
to distinguish between countable and uncount-
able compound nouns. 
θ≥
),(
),(
isNf
areNsf
     (3) 
The model 3 that compares the frequencies 
of noun-be pairs (eg,. f(“account books are”), 
f(“account book is”) is used to distinguish plural 
only and countable compound nouns. 
With the help of WWW search engine 
Google, the frequencies (web hits) in the models 
can be obtained using quoted n-gram queries 
(“soft surroundings”). Although in Keller and 
Lapata (2002), they experimentally showed that 
web-based approach can overcome data sparse-
ness for bigrams, but the problem still exists in 
our experiments. When the number of pages 
found is zero, we smooth zero hits by adding 
them to 0.01. 
Countable compound nouns create some 
problems when we need to pluralize them. For 
no real rules exist for how to pluralize all the 
words, we summarized from “Guide to English 
Grammar and Writing” for some trends. We 
processed our experimental data following the 
rules below.  
1. Pluralize the last word of the compound 
noun. Eg,. bedrooms, film stars. 
2. When “woman” or “man” are the modi-
fiers in the compound noun, pluralize 
both of the words. Eg,. Women-drivers. 
3. When the compound noun is made up as 
“noun + preposition (or prep. phrase)”, 
pluralize the noun. Eg,. fathers-in-law. 
4. When the compound noun is made up as 
“verb (or past participle) + adverb”, plu-
ralize the last word. Eg,. grown-ups, 
stand-bys. 
Although the rules cannot adapt for each 
compound noun, in our experimental data, all 
the countable compound nouns follow the rules. 
We are sure that the rules are viable for most 
countable compound nouns. 
Although we used Google as our search en-
gine, we did not use Google Web API service 
for programme realization, for Google limits to 
1000 automated queries per day. As we just 
need web hits returned for each search query, we 
extracted the numbers of hits from the web 
pages found directly.  
3 Experimental Data 
The main experimental data is from Webster’s 
New International Dictionary (Second Edition). 
The list of compound words of the dictionary is 
available in the Internet
2
. We selected the com-
pound words randomly from the list and keep 
the nouns, for the word list also mixes com-
pound verbs and adjectives with nouns together. 
We repeated the process several times until got 
our experimental data. We collected 3000 words 
for the training which is prepared for optimizing 
the detecting thresholds, and 500 words for the 
test set which is used to evaluate our approach. 
In the sets we added 180 newcoined compound 
nouns (150 for training; 30 for test). These rela-
tively new words that were created over the past 
seven years have not yet reached any dictionari-
es
3
.  
 
Countability Training set Test set 
Plural only 80 21 
Countable 2154 352 
Uncountable 766 127 
Total 3000 500 
Table 1. The make-up of the experimental data  
We manually annotated the countability of 
these compound nouns, plural only, countable, 
uncountable. An English teacher who is a native 
speaker has checked and corrected the annota-
tions.  The make-up of the experimental data is 
listed in Table 1. 
                                                           
2
 The compound word list is available from http://www. 
puzzlers.org/wordlists/dictinfo.php. 
3
 The new words used in the paper can be found in http: 
//www.worldwidewords.org/genindex-pz.htm 
105
4 Experiments and Results 
4.1 Detecting plural only compound nouns 
Plural only compound nouns that have not sin-
gular forms always occur in plural forms. The 
frequency of their singular forms should be zero. 
Considering the noise data introduced by search 
engine, we used model (1) and (3) in turn to de-
tect plural noun. We detected plural only com-
pound nouns with the following algorithm 
(Figure 2), which is used to distinguish between 
plural only and non-plural only compound. 
 
 
 
 
 
 
 
 
 
 
Figure 2. Detecting algorithm for plural only  
 The problem is how to decide the two 
thresholds. We preformed exhaustive search to 
adjust θ 1,θ 2 optimized on the training set. 
With 0  ≤ θ 1,θ 2 20, all possible pair values 
are tried with the stepsize of 1. 
≤
AB
A
call =Re                  (4) 
AC
A
ecision =Pr             (5) 
callecision
callecision
scoreF
RePr
RePr2
+
××
=−  (6) 
We use Recall and Precision to evaluate the 
performance with the different threshold pairs. 
The fundamental Recall/Precision definition is 
adapted to IE system evaluation. We borrowed 
the measures using the following definition for 
our evaluation. For one experiment with a cer-
tain threshold pair, A stands for the number of 
plural found correctly; AB stands for the total 
number of plural only compound nouns in train-
ing set (80 words); AC stands for the total num-
ber of compound nouns found. The Recall and 
Precision are defined in (4) and (5). We also 
introduced F-score when we need consider the 
Recall and Precision at the same time, and in the 
paper, F-score is calculated according to (6). 
Figure 3 shows the performance  evaluated 
by the three measures when θ 1=8 and 0 ≤ 
θ 2≤10 with a stepsize of 1. We set θ 2 to 5 for 
the test later, and accordingly the values of Re-
call, Precision and F-score are 91.25%, 82.95% 
and 87.40% respectively. 
Recall/Precision/F-score
0%
20%
40%
60%
80%
100%
012345678910
threshold
Accuracy
Recall
Precision
F-score
Figure 3. The Recall/Precision/F-score graph 
(θ 1=8 and 0≤ θ 2≤10) 
if  ( 1
)(
)(
θ≥
Nf
Nsf
) 
    then plural only; 
else if ( 2
),(
),(
θ≥
isNf
areNsf
)  
then plural only; 
else  
countable or uncountable; 
4.2 Detecting uncountable compound 
nouns 
Uncountable compound nouns that have not plu-
ral form always occur in singular form.  
 
 
if ( N is not plural only) 
 
then if  ( 3
)(
)(
θ≥
Nsf
Nf
)   
 
then uncountable; 
 
else if ( 4
),(
),(
θ≥
Nsmanyf
Nmuchf
) 
 
 
 
   then uncountable; 
else  
countable;  
 
 
 
Figure 4. Detecting algorithm for uncountable 
compound nouns 
The algorithm detecting uncountable compound 
nouns is shown in Figure 4. Using model (1) and 
(2), we attempted to fully make use of the char-
acteristic of uncountable compound nouns, that 
is the frequencies of their occurrence in the sin-
gular forms are much larger than that in the plu-
ral forms.  
106
The method to obtain the optimal threshold 
θ 3 and θ 4 is the same to 4.1. We set θ 3 to 24, 
θ 4 to 2, and the values of Recall, Precision and 
F-score are 88.38%, 80.27% and 84.13% respec-
tively. 
4.3 Performance on the test suite 
We evaluated our complete algorithm with the 
four thresholds (θ 1=8,θ 2=5, θ 3=24, θ 4=2) 
on the test set, and the detecting results are 
summarized in Table 2. There are 352 countable 
compound nouns in our test set, then when clas-
sify all the test words as countable, we can at 
least get the accuracy of 70.4%. We used it as 
our baseline. The accuracy on the total test date 
is 89.2% that significantly outperforms the base-
line. For the 30 newcoined compound nouns, the 
detecting accuracy is 100%. This can be ex-
plained by their infrequence. Newcoined words 
are not prone to produce noise data than others 
just because they are not occurring regularly.  
 
 Correct Incorrect Recall Precision F-score 
Plural only 18 4 85.71% 81.81% 83.71% 
Countable 320 22 90.90% 93.57% 92.22% 
Uncountable 108 28 85.04% 79.41% 82.15% 
Total 446 54 89.2% 89.2% 89.2% 
Table 2. The accuracy on the test suit
5 Conclusion 
From the results, we show that simple 
unsupervised web-based models can achieve the 
promising results on the test data. For we 
roughly adjusted the threshold with stepsize of 1, 
better performance is expected with stepsize of 
such as 0.1.  
It is unreasonable to compare the detecting 
results of individual and compound nouns with 
each other since using web-based models, com-
pound nouns made up of two or more words are 
more likely to be affected by data sparseness, 
while individual nouns are prone to produce 
more noise data because of their high occur-
rence frequencies. 
Anyway using WWW is an exciting direc-
tion for NLP, how to eliminate noise data is the 
key to improve web-based methods. Our next 
step is aiming at evaluating the internet resource, 
distinguishing the useful and noise data. 
References 
Baldwin, Timothy and Francis Bond.2003. Learning 
the countability of English nouns from corpus data. 
In Proceedings of the 41
st
 Annual Meeting of the 
Association for Computational Linguistics. Sap-
poro, Japan, 463-470. 
Francis Bond and Caitlin Vatikiotis-Basteson. 202. 
Using an ontology to determine English countabil-
ity. In Proceeding of the 19
th
 International confer-
ence on computational Linguistics (COLING 
2002), Taipei, Taiwan. 
Keller, F, Lapata, M. and Ourioupina, O. 2002. Us-
ing the web to overcome data sparseness. In Pro-
ceedings of the Conference on Empirical Methods 
in Natural Language Processing. Philadelphia. 
230-237.  
Keller, F and Lapata, M. 2003. Using the web to ob-
tain frequencies for unseen bigrams. Computa-
tional Linguistics 29, 3, 459-484. 
Lapata, M and Keller, F. 2004. The web as a baseline: 
Evaluating the performance of unsupervised web-
based models for a range of NLP tasks. In Pro-
ceedings of the Human Language Technology 
Conference of the North American Chapter of the 
Association for Computational Linguistics. Boston,  
Lapata, M and Keller, F. Web-based Models for 
Natural Language Processing. To appear in 2005 
ACM Transactions on Speech and Language 
Processing. 
Lane O.B. Schwartz. 2002. Corpus-based acquisition 
of head noun countability features. Master’s thesis, 
Cambridge University, Cambridge, UK. 
Guide to English Grammar and Writing. http:// cctc. 
commnet.edu/grammar/
107
