NEW METHOD USING DECLINABLE WORDS AND
CONCURRENT WORDS TO CREATE A LARGE NUMBER OF
FA WORDS
El-Sayed Atlam, K. Morita, M. Fuketa and Jun-ichi Aoe
Department of Information Science and Intelligent Systems, University of Tokushima, Tokushima, 770-8506, Japan
Keywords: FA Words, Declinable words, Concurrent Words, CFA Words, Recall, Precision.
Abstract: The Readers can know the subject of many document fields by reading only some specific Field Association
(FA) words. Document fields can be decided efficiently if there are many rank 1 FA words (words that
direct connect to terminal fields) and if the frequency rate is high. This paper proposes a new method for
increasing rank 1 FA words using declinable words and concurrent words which relate to narrow
association categories and eliminate FA word ambiguity. Concurrent words become Concurrent Field
Association Words (CFA words) if there is a little field overlap. Usually, efficient CFA words are difficult to
extract using only frequency, so this paper proposes weighting according to degree of importance of
concurrent words. The new weighting method causes Precision and Recall to be higher than by using
frequency alone. Moreover, combining CFA words with FA words allow easy search of fields which can not
be searched by using only FA words.
1 INTRODUCTION
Determining keywords is crucial in successful
Information Retrieval (IR). After automatic
classification of document fields (
Fukumoto, 1996),
Vector Model (Iwayama, 1999
) and Probabilistic
Model (Fuhr, 1989) can use general information
about document files to calculate degree of
document similarity. However, there are problems
because of multiple topics and collections of
document fields, and so content to be searched
usually exists only in part of the file (Callen , 1994;
Melucii, 1998).
In this paper, field means basic and common
knowledge used in human communication (Tsuji et
al., 1999). Readers know topic super-field (e.g.
Sports) or sub-field (e.g. Baseball) of document
fields based on specific Field Association (FA)
words in that document. For example, the word
“election” can indicate super-field <Politics> and the
word “home run” can indicate sub-field <Baseball>.
In this paper, FA words are ranked according to
document field. Rank 1 FA words are relatively few
and can be used efficiently to decide document
fields. FA words in other ranks are always numerous
and are not so helpful for deciding document fields.
Document fields can be decided easily if there are
many rank 1 FA words and frequency rate is high.
Document fields can not be decided easily if there
are few rank 1 FA words or if the FA words appear
in overlapping document fields.
To overcome problems associated with rank 1
FA words, this paper proposes a new method using
declinable words and concurrent words to create a
relatively large number of rank 1 FA words and to
eliminate ambiguous FA words.
a) Declinable words express action or condition.
To eliminate ambiguity of the FA word, declinable
words are combining with FA words. For example,
FA word “pass” is ambiguous and associated with
many sub-fields of <SPORTS>, but combining
declinable word “through” with “pass” creates
“through-pass” which associate with sub-field
<Soccer>.
b) Concurrent words (C words) usually have two
short unit FA words connected by particles (e.g. the,
in, and) and short unit information can be used to
associate with fields. The weight function of C
words can be expressed by the weight function of
the short unit FA words.
527
Atlam E., Morita K., Fuketa M. and Aoe J. (2010).
NEW METHOD USING DECLINABLE WORDS AND CONCURRENT WORDS TO CREATE A LARGE NUMBER OF FA WORDS.
In Proceedings of the 2nd International Conference on Agents and Artificial Intelligence - Artificial Intelligence, pages 527-531
DOI: 10.5220/0002709705270531
Copyright
c
 SciTePress

2 FIELD ASSOCIATION WORDS
Field Association (FA) words can identify
documents. Short unit association words are
minimum meaningful units which can not be divided
without loss of meaning (Atlam, 2002; Atlam 2006).
For example, “pitcher” and “home run” are short
unit association words that can be associated with
terminal field <Baseball>.
2.1 Document Field Tree
A document field tree structure represents
relationships between ranked document fields (Aoe,
1989;
Breiman 1994; Elmarhomy 2006; Salton 1983;
Salton 1988). In field tree structure, a leaf node is a
terminal document field and other nodes are middle
document fields. In this study, based on Imidas’99
(Dozawa, 1999) term dictionary, the field tree
contains 14 main (parent) fields, 18 middle fields
and 172 terminal (child) fields.
2.2 Ranking FA Words
FA words extracted from Corpus data may have
various association field ranks. Some FA words may
associate with only one terminal field or one middle
field; other FA words may associate with several
terminal fields or several middle fields. FA word w
may be defined according to five ranks:
Rank 1: Complete FA word w associates with only
one terminal field.
Rank 2: Quasi complete FA word w associates with
a limited number of terminal fields which have the
same parent field (main-field).
Rank 3: Middle FA w associates with only one
middle field.
Rank4: Intersection FA word w associates with
several middle fields or several fields.
Rank 5: Non association- word w does not associate
with any specific field.
2.3 Constructing FA Words
2.3.1 Basic Outline
To construct FA words, it is first necessary to extract
candidate FA words from Corpus files classified
manually into related fields. Table 1 shows some
extracted candidate FA words providing information
such as: candidates (A) are extracted from field (B)
at a frequency (x times).
2.3.2 FA Words and Declinable Words
To eliminate ambiguity, the present method
combines FA words with declinable words from
Corpus documents which are classified beforehand
by Tree structure.
In Table 1, “pass” is an FA word for
<Baseball>, <Soccer>, and <Basketball>. However,
in this document data the action “through-pass” only
exists in <Soccer >, so “through-pass” is considered
to be an FA word for <Soccer>. “Game” is mainly
an FA word used in middle field <Ball Game>.
However, the action “a perfect game” is only in
<Baseball >, so “a perfect game” is considered to be
an FA word for <Baseball>. “recommendation” is
an FA word for fields <Election> and <Entrance
Examination>,
Table 1: Sample of FA Words/ Declinable Words and
Association Fields with Ranks.
FA Words Field Rank
Declinabl
e words
Filed Rank
pass
<Baseball>
<Soccer>,
<Basketball
>
2
through-
pass
< Soccer > 1
game <SPORTS> 3
a perfect
game
<Baseball > 1
Recomme-
ndation
<Election>,
<Entrance
Examination
>
4
recommend
ation
recruitment
<Entrance
Examination>
1
fish
No
Association
Field
--
extinction
of fish
<Environment
Problem>
1
but “recommendation recruitment” only exists in
<Entrance Examination>, so “recommendation
recruitment” is an FA word for <Entrance
Examination>. “Fish” is not an FA word because
“fish” can not be used to associate with any field.
But the condition “Extinction of fish” is an idiomatic
expression used only in a specific field, and so
“Extinction of fish” is an FA word for <Environment
Problems>.
For all FA words/declinable words, fields
can not be necessarily specified. For example,
“strike”, “hit” and “shoot” have different meaning
in <Baseball>, <Soccer> and <Basketball>.
Combining “strike” or “hit” with “shoot” does not
produce an expression which can be used in either of
the three fields. Generally, association fields are not
necessary identified by combining FA words with
declinable words. However, the range of association
fields can be limited by pairing FA words having
meaningful relationship.
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
528

3 CONCURRENT WORDS AND
ATTACHING WEIGHT
3.1 Concurrent Words
Concurrent words (C words) are two short unit FA
words connected by particles (e.g. the, in, and)
which are used to associate fields. The importance of
C words can be expressed by ranking the weight of
the short unit FA words. The importance of C words
relates especially to appearance frequency and to
association fields of the short unit FA words. The
frequency of short unit words shows field rank, and
number of overlapping fields shows the degree of
ambiguity of the short unit words.
In this paper, it is assumed that no rank 1 short
unit FA words are C words because rank 1 FA words
refer to specific fields and it is not necessary to
converge association fields.
3.1.1 Attaching Weight
Generally, to extract a word which characterizes a
file, a weight function TF x IDF attaches to the
words (TF is a high frequency of the appearance
characteristic words and IDF is inverse document
Frequency). However, not every word with high
frequency characterizes a file. For example, particles
(the, to, etc) appear often in a file, but the particles
are not characteristic words. On the other hand,
some characteristic words have relatively low
frequency, so IDF attaches high weight to those
characteristic words and considers weight in many
fields. IDF value is given by log N/df(t), where total
number of files is N and the number of files which
include word t is df(t). TF x IDF is given by:
W(d,t) = TF (d,t) x IDF (t)**************(1)
where TF is the normalized frequency value of a
word t in a file d.
This research applies TF x IDF to consider
the normalized frequency of a word
α in one field A.
So, the weight of a short unit word
α can be defined:
N
Weight
A
( α ) = Freq
A
( α ) x log ( ) (2)
Category _ num (
α )
where Freq is the normalized frequency of word
α in field A, N is total number of fields and
Category_num is number of fields containing
α .
In the same way, the weight of word
β in Field
A can be calculated:
N
Weight
A
( β ) = Freq
A
( β ) x log ( )
Category _ num (
β )
Consider a C word
α + β is in a field A, the weight
of the C words is:
Weight
A
( α + β ) = Weight
A
( α ) + Weight
A
( β ) =
(3)
N N
Freq
A
( α ) x log ( ) + Freq
A
( β ) x log( )
Category _ num (
α ) Category _ num ( β )
The following cases are examples of weight
according to degree of importance of C words:
Case (1): C words with high frequency are
confirmed to be improper for use as CFA words.
In field <Soccer>
Foreigner Freq. = 52 Category_num (foreigner)= 35
athlete Freq. = 535 Category _ num (athlete) = 57
foreigner and athlete Freq. = 52
Cross_Category_num = 26
foreigner and athlete (frequency rank)= 13
foreigner and athlete (weighting according to
degree of importance rank) = 408
W
new
(foreigner and athlete) =
52 x log (133/35) + 535 x log (133/57)
52 x = 58.27 26
In field <Soccer>, the concurrent relation of
“foreigner” and “athlete” has frequency of 52. If C
words are ranked according to frequency, provide
relatively high rank of 13 in field <Soccer>. So, C
words might appear to be important by considering
only frequency, but the concurrent relation of
“foreigner” and “athlete” is not characteristic words
in field <Soccer>; “foreigner” and “athlete” appear
in all sub- fields of field <SPORTS>.
Ranking “foreigner” and “athlete” by weighting
according to degree of importance provides a
relatively low rank of 408. So, C words “foreigner”
and “athlete” are not CFA words in field <Soccer>.
4 EVALUATION RESULTS
4.1 Field Systems and Test Data
To verify the efficiency of the new method
described in this paper, about 38,000 articles from a
data set of 20 Newsgroups from CNN Web Site
(1995-2001) were selected. There were various
topics related to sports, computers, politics,
economics, etc. The accumulating method is to
search titles of articles by using keywords exists in
field tree system.
4.2 Method Evaluation
Precision and Recall are evaluated to show how well
weighting according to degree of importance
NEW METHOD USING DECLINABLE WORDS AND CONCURRENT WORDS TO CREATE A LARGE NUMBER
OF FA WORDS
529

calculated by the new method expresses CFA words
in specific fields. C words with higher weighting
according to degree of importance are judged to
determine CFA words. The highest 10%, 20%, and
30% ranking of weight according to degree of
importance is used to calculate Precision and Recall.
Efficiency of this method is also estimated.
The test is done by the following sequence:
Step 1: C words are classified manually according
to the possibility of field association.
Step 2: Weighting according to degree of
importance is attached to C words.
Step 3: Precision (P) and Recall (R) are evaluated as
Number of CFA words in the extracted C words
Precision (P) = ---------------------------------------
Total number of C words automatically extracted
Number of CFA words in the extracted C words
Recall (R)= -------------------------------------
Total Number of CFA words manually extracted
Tables 4 and 5 show P and R of C words in
field <Soccer> are arranged by weighting according
to degree of importance and frequency.
Table 3: Sample C words Arranged by the Degree of.
C words Degree o
f
Importance
Frequenc
y
CFA word
1- election area 8703.733 450 Yes
2- election delegate 8193.303 251 Yes
3- vote counting 941.477 123 Yes
4- government election 479.294 28 Yes
5- class being busy 433.257 17 NO
6- seat obtain 284.514 27 Yes
7- proportion turnout 221.666 2 Yes
8- wide support 193.435 16 No
9- prefecture diet time 159.615 6 No
10- seat assure 124.5 12 Yes
Table 4: Precision & Recall in Field <Soccer> by
Weighting According to Degree of Importance .
Rank P(%) R(%)
Upper 10% 79 71
Upper 20% 48 87
Upper 30% 34 93
Upper 40% 26.4 95.7
Upper 50% 21.3 96.7
Upper 60% 18 98.3
Upper 70% 15.6 99
Table 5: Precision & Recall in Field <Soccer> According
to Frequency.
Rank P(%) R(%)
Upper 10% 28 25
Upper 20% 22 40
Upper 30% 30 81
Upper 40% 25 93
Upper 50% 21 96
Upper 60% 18 97
Upper 70% 15 99
5 CONCLUSIONS
Document fields can be decided efficiently if there
are many rank 1 FA words and if the frequency rate
is high, but generally, there are limited rank 1 FA
words, especially when there are few Corpus
documents. This paper proposes a method for
deciding FA words using C words and declinable
words which relate to narrow association categories
and eliminate FA word ambiguity.
Usually, efficient CFA words are difficult to
extract using frequency only. This paper proposes a
new efficient method for weighting according to
degree of importance of C words, causing P and R to
be higher than by using frequency alone. R and P
significantly increase by using C words ranked in
the top 10% weighted according to degree of
importance. R and P decrease somewhat when C
words are ranked between 10% to 50% by weighting
according to degree of importance because there are
many ambiguous words. Future research could focus
on clustering C words and FA words.
REFERENCES
Aoe, J., Morita K., and Mochizuki H. 1989. An Efficient
Retrieval Algorithm of Collocate Information Using
Tree Structure. Transaction of The IPSJ, 39 (9),
pp.2563-2571.
Atlam, E.-S., Morita K., and Aoe, J. 2002. A New Method
For Selecting English Compound Terms and its
Knowledge Representation. Information P. &
Management Journal, Vol. 38, No. 6, pp. 807-821.
Atlam, E.-S., Morita, K., Fuketa M., and Aoe, J. 2006.
Automatic Building of New Field Association Word
Candidates Using Search Engine”, Information
Processing & Management Journal, Vol.42, No. 4,
pp.951-962.
Breiman, L., Friedman, J.H., Olshen R. A. and Stone C.J.
1994. Classification and Regression Trees. Chapman
Hall.
Callen, J. P. 1994. Passage and level evidence in
document retrieval. In Proc. of the 17th Annual
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
530

International ACM SIGIR Conference on Research
and Development in Information Retrieval, pp.302-
310
Dozawa, T. 1999. Innovative Multi Information
Dictionary Imidas’99, Annual Series, Zueisha
Publication Co., Japan ( In Japanese).
Fuhr, N. 1989. Models for retrieval with probabilistic
indexing, Information Processing and Retrieval, 25
(1), 55-72.
Elmarhomy, G., Atlam, E.-S., Morita, K., Fuketa, M. and
Aoe, J. 2006. Automatic Deletion of Unnecessary
Field Association Word Using Morphological
Analysis, International Journal of Computer and
Mathematics, Vol. 83, 3, pp 247-262.
Fukumoto, F., Suzuki, Y. 1996. Automatic Clustering of
Articles using Dictionary definitions. Proceeding of
the 16
th
International Conference on Computional
Linguistic (COLING’96), 406-411.
Iwayama, M. and Tokunaga, T., 1999. Probabilistic
Passage Categorization and Its Application. Journal of
Natural language Processing. Vol. 6 No. 3, pp. 181-
198.
Melucii, M. 1998. Passage Retrieval and a Probabilistic
technique. Information Processing and Management,
34(1), 43-68.
Salton G., and McGill M.J., 1983. Introduction of Modern
Information Retrieval. New York: McGraw-Hill.
G. Salton. 1988, Automatic Text Processing: The
Transformation, Analysis, and Retrieval of
Information by Computer. Addison-Wesley.
Tsuji, T., Nigazawa, H., Okada, M. and Aoe, J. 1999.
Early Field Recognition by Using Field Association
Words. In Conference on Computer Processing of
Oriental Language, Vol. 2, pp. 301-304.
NEW METHOD USING DECLINABLE WORDS AND CONCURRENT WORDS TO CREATE A LARGE NUMBER
OF FA WORDS
531