Microsoft Word - Hierarchies 717 for dke 14sep04.doc


S. Levachkine & A Guzman-Arenas  1 of 31 

Hierarchy as a new data type for qualitative variables 
 

Serguei Levachkine, Adolfo Guzman-Arenas* 
Centro de Investigación en Computación, Instituto Politécnico Nacional 

“Lopez Mateos” Campus, Mexico City, MEXICO 
sergei@cic.ipn.mx, aguzman@ieee.org 

 
SUMMARY. Qualitative variables take symbolic values such as cat, orange, California, 

Africa. Often these values can be arranged in levels of deeper detail. For example, the vari-

able place_of_birth takes as level-1 values Africa, Asia... as level-2 values Nigeria, Japan... 

as level-3 values California, Massachusetts... These values are organized in a hierarchy H, 

a mathematical construct among these values. Over H, the following are defined: (1) the 

function confusion resulting when using a symbolic value instead of another; (2) the close-

ness to which object o fulfills predicate P; (3) a method which allows precision-controlled 

retrieval for relational databases whose objects have symbolic values. 

 

INDEX TERMS: hierarchy, ontology, approximate queries, confusion, knowledge represen-

tation. 

1. Introduction 

What is the capital of Germany? Berlin is the correct answer; Frankfurt is a close miss, 

Madrid a fair error, and sausage a gross error. What is closer to a cat, a dog or an orange? 

Can we measure these errors and similarities? Can we retrieve objects in a data base that 

are close to a desired item? Yes, by arranging these symbolic (that is, non-numeric) values 

in a hierarchy. For the sake of completeness, four different definitions of hierarchy are 

                                                           
* Corresponding author. 



S. Levachkine & A Guzman-Arenas  2 of 31 

given in §2. At least one of them is original. These definitions are important for under-

standing what a hierarchy is. However, the confusion does not depend on particular defini-

tion.  

1. This arrangement allows the definition of confusion (§3.1) to measure the error when 

using one symbolic value associated with a node in a hierarchy in place of the (in-

tended, correct) symbolic value associated with other node in the hierarchy. Variations 

of the definition are: 

a. When the values represent sets with some size, as population in {France, Italy, 

Spain, Sweden} we talk about percentage hierarchies (§3.1.1). 

b. When the values can be ordered, as temperature in {frigid, cold, warm, hot, 

burning} we talk about ordered hierarchies (§3.1.2). 

2. Confusion is also defined (§3.3) for hierarchies whose nodes are associated with predi-

cates (called variables in the paper) in addition to hierarchies whose nodes are associ-

ated with values of a variable (item 1). 

3. Confusion among values (item 1) and among variables (item 2) enables measurement of 

how close a given object o fulfils a given predicate P (§3.2.2), and we write Pε(o) for 

this measure. 

The main contributions are (1)-(3), pertaining to symbolic values. Of course, errors, dis-

tances and approximate answers are well understood and developed for numerical values. 

The rest of the introduction discusses related work. Section 2 gives several definitions 

for hierarchies. Section 3.1 introduces confusion. Section 3.2 presents predicates on hierar-

chies. Discussion of overall paper’s results is in Section 4 and conclusions form Section 5. 

 



S. Levachkine & A Guzman-Arenas  3 of 31 

Related work. Artificial Intelligence, Natural Language and Knowledge Representation 

communities have been gauging the distance, proximity or “relatedness” between symbolic 

values. Relevant efforts: 

A. Hierarchies. The concept of a (generalization) hierarchy is not new. Hierarchies are 

used in data warehousing and data mining; see, for instance, the H-sets of Bhin [2]. A 

practical use of hierarchies in symbolic processing is Clasitex [9], which finds the 

themes of an article written in Spanish or English. It uses the concept tree, and a word 

(not in the tree) suggests the topic of one or more concepts in the tree. BiblioDigital© 

[4], a recent development, uses a large taxonomy (although not a hierarchy) to classify 

text documents; a (distributed) crawler in it retrieves “external” documents residing 

elsewhere in the Web. If a document is about (Cf. Clasitex) war, Iraq and President 

Bush, its URL will be stored in these three nodes in the concept tree. Hierarchies are 

simpler than ontologies, albeit very useful [13, 21].  

The data modeling community, through the entity-relationship model, also organize 

items by their nature, properties and the relations among them. 

B. Natural Language. Linguists (see, for instance, Proceedings of CICLING 04, LNCS 

2945, as referenced in [11]) have proposed many versions of semantic closeness, simi-

larity, and other measures among words. Everett [6] identifies conceptually similar 

documents using a single ontology. Sidorov [8] does the same using a topic hierarchy: a 

kind of ontology. Montes y Gómez [19] builds trees of words, and by graph matching 

retrieves similar texts. Another common idea twisting around is to regard the represen-

tation space with a “universal” measure of proximity of space’s elements and then an 

attempt to adapt it to different subject domains [16] [24]. Comments on this in §4. 



S. Levachkine & A Guzman-Arenas  4 of 31 

WordNet [26] organizes information in logical groupings called synsets; each syn-

set is a list of synonymous words or collocations (e.g., “fountain pen”, “take in”), and 

pointers that describe the relations between this synset and other synsets. A word or 

collocation may appear in more than one synset, and in more than one part of speech. 

The words in a synset are logically grouped such that they are interchangeable in some 

context. Nouns and verbs are organized into hierarchies based on the hy-

pernymy/hyponymy relation between synsets. Two kinds of relations are represented by 

pointers: lexical and semantic. Lexical relations hold between word forms; semantic re-

lations hold between word meanings. These relations include (but are not limited to), 

antonymy, entailment, and meronymy/holonymy. Additional pointers are used to indi-

cate other relations. 

Budanitsky [3] compares five measures of similarity or semantic distance in Word-

Net: Jiang and Conrath's measure (the best in the comparison: a spelling-corrector on 

real data); that of Hirst-St-Onge (seriously over-related), that of Resnik (seriously un-

der-related [23]), and those of Lin [16] and of Leacock-Chodorow (in between). Note 

that all the measures except those of Hirst and St-Onge are similarity (not relatedness) 

measures considering only the hyponymy hierarchy of WordNet. The main problem 

(§4) with these approaches is that they use distances, thus obeying the symmetric prop-

erty d(a,b) = d(b,a), while conf (§3.1) does not. 

C. Ontologies. At least three approaches appear when measuring similarity or relatedness 

of concepts (nodes in the ontology): 



S. Levachkine & A Guzman-Arenas  5 of 31 

1. Syntactic approach. Methods that take into account only the organization of the 

tree or data structure of the ontology; for instance [19], those based on XML, or 

the “ontology merging” of Protégè [20]. 

2. Standard ontology. Use of a common or agreed-upon ontology. Clearly, if dif-

ferent people (or agents) use the same ontology, similarities among concepts 

will be consistently measured across users. CYC [7] was an early attempt to 

build the concept tree for common concepts.  A common ontology is predicted 

in [11]; conceptually similar documents are identified in [6] by using a single 

ontology. In contrast, point (3) following shows use of different ontologies. 

3. Measuring similarity across ontologies. LIA, a language for agent interaction 

[10, 13, 21], has an ontology comparator COM, that maps a concept from one 

ontology into the closest corresponding concept in another ontology. COM is 

used in sim of §3.4. By repeated use of sim, the degree of understanding du(B, 

OA) of agent B (with ontology OB) about ontology OA is found in [22]. 

Instead of using ontologies, this paper works on arbitrary hierarchies (§2). Why? Be-

cause the problem-oriented interaction can be easier to maintain if the hierarchical 

structure is not a priori rigid as in the case of common hierarchies or ontologies. 

D. Pattern Classifiers. Our predicates with controlled precision or confusion (§3.2.1) are 

similar to Pattern Classifiers [18], but these classify objects according to the values of 

their properties, whereas hierarchies help to classify these values, when they are non-

numeric. 

E. Distances and ultradistances. Traditionally [1, 25], the representation space is re-

garded as a metric space with some “exotic” distance (e.g., ultrametric distance to 



S. Levachkine & A Guzman-Arenas  6 of 31 

measure the “distances” between members of a hierarchy). Thus, §2.4 develops ul-

trametric distances for hierarchies. However, often is not the case that such a distance 

meets the needs of the classification problem under consideration. Thus, we lean to-

wards functions like conf (§3.1) that are not distances. 

2. Theory 

This section continues with the focus on distances of item (E) of §1: we show how to build 

an ultradistance from a hierarchy (§2.4.1), how to build a hierarchy from an ultradistance 

(§2.4.2), whereas in Section 3 we move to a new approach that does not use distances. 

Element set E. A set whose elements are explicitly defined. ♦ 1 Example: {red, blue, white, 

black, pale}. 

Ordered set. An element set whose values are ordered by a < (“less than”) relation. ♦  Ex-

ample: {very_cold, cold, warm, hot, very_hot}.  

Covering. K is a covering for set E if K is a set of subsets si ⊂  E, such that ∪  si = E. ♦  

Every element of E is in some subset si ∈  K. If K is not a covering of E, we can make it 

so by adding a new sj to it, named “others”, that contains all other elements of E that do 

not belong to any of the previous si. 

Exclusive set. K is an exclusive set if si ∩ sj = ∅ , for every si, sj ∈  K. ♦  Its elements are 

mutually exclusive. If K is not an exclusive set, we can make it so by replacing every 

two overlapping si, sj ∈  K with three: si - sj, sj - si, and si ∩ sj. 

Partition. K is a partition of set E if it is both a covering for E and an exclusive set. ♦  

                                                           
1
 This symbol means: end of definition. 



S. Levachkine & A Guzman-Arenas  7 of 31 

Symbolic value. A value that is not numerical, vector or quantitative. ♦  Example: red. 

Representation. A symbolic value v represents a set E, written v ∝  E, if v can be consid-

ered a name or a depiction of E. ♦  v is associated with E. Example: strings ∝  {violin, 

viola, cello, guitar}. 

Qualitative variable. A single-valued variable that takes symbolic values. ♦  Its value can-

not be a set,2 although such value may represent a set. 

father_of (v). In a tree, f is the father_of v if f is the node immediately  following v in the 

path from v to the root. ♦  f is “the node from which v hangs.” We say that v is a son_of 

(f). ♦  Similarly, grand_father_of (v), brothers_of, aunt, ascendants, descendants... 

are defined, when they exist. ♦  The root is the only node that has no father. 

2.1 Hierarchy 

Definition 1. A hierarchy H of an element set E is a tree whose root is E and if a node has 

sons then these form a partition of their father. ♦  

Definition 2. For an element set E, a hierarchy H of E is a tree of nodes; each node n is 

either an element of E or a set of symbolic values vi, for i=1,…n, where vi ∝ Ei, and {E1, 

E2,…, En} is a partition of E. ♦  Example: for E = {chair table bed shirt loafer moccasin 

hammer paintbrush broom saw}, a hierarchy is (figure 1) H1 = {furniture ∝  {chair table 

bed} apparel ∝  {shirt shoe ∝ {loafer moccasin}} tool ∝  {hammer brush ∝  {paintbrush 

broom} saw}}. A hierarchy groups E into smaller sets of alike symbolic values. 

                                                           
2
 Variable, attribute and property are used interchangeably. Some objects have an attribute (such as weight) while 

others do not: the weight of blue does not make sense, does not exist. A variable (color, height) describes an aspect of 
an object; its value (blue, 2 Kg) is such description (symbolic value) or measurement (numeric value). 



S. Levachkine & A Guzman-Arenas  8 of 31 

merchandise 

furniture 

hammer 

tool 

chair table bed shirt shoe

apparel 

brush saw 

loafer moccasin paintbrush broom 

Definition 1 emphasizes that the nodes of H are sets, while for definition 2 the nodes are 

symbolic values (such as furniture). To reconcile, we use the former definition v ∝  E, 

where a symbolic value v represents a set E. 

Some times, we add node “others” to certain level of a hierarchy, when we are not sure 

whether the nodes already present at that level will collectively exhaust their father. Thus, 

for instance, in Figure 1, we could add to the second level node “other_merchandise”, if we 

are not sure that furniture, apparel and tool comprise all the merchandise we are interested. 

A hierarchical variable is a single-valued qualitative variable whose values belong to a 

hierarchy. ♦  The data type of a hierarchical variable is hierarchy. Example: trades_in 

that takes values from H1, as trades_in = furniture, trades_in = broom. 

 

 

 

 

 

 

Fig. 1. Hierarchy H1 of articles for sale. 

2.2 Partitions of a finite set 

Let E be a set of n elements. A partition P of E is a set of k subsets Ci of E such that 

(1) Ci∩Cj=∅ ; (2) ∪ iCi=E. ♦  

Two elements x and y of E are equivalent in a partition P if they belong to the same 

class Ci; this is denoted by xPy. ♦  



S. Levachkine & A Guzman-Arenas  9 of 31 

Let P(E) be the set of all partitions of E; an order relation among the members of 

P(E), denoted by <, can be defined thus: for any two partitions P and P’, P<P’ iff 

xPy→xP’y. Partition P is said to be finer than P’; it has more classes than P’, i.e. k>k’. ♦  

Example: let E={ a,b,c,d,e,f} . Then E is less fine (i.e. coarser) than {{ a} ,{ b,c,d} ,{ e,f}}  

which in turn is less fine than {{ a} ,{ b,c} ,{ d} ,{ e} ,{ f}} . 

A lattice structure for P(E) can be based on the order relation. For every pair of parti-

tions P and P’ there is a least upper bound (l.u.b.) P∨ P’, and a greatest lower bound (g.l.b.) 

P∧ P’. ♦  

Let us call Pk a partition of k classes where k is the level of Pk. A partition P’ is said to 

cover a partition P if and only if P’ results from combining two classes of P. ♦  Note that 

P’=bcd,a does not cover P=ab,c,d, because P’ cannot be obtained from the union of two 

classes of P, which would in fact give P’1=abc,d, P’2=abd,c and P’3=ab,cd, but not P. 

A chain in the lattice is a sequence of partitions in order, e.g. (P1, P2,...,Pj) where P1<P2 

<...< Pj; the term is understood in the sense of an elementary chain in graph theory. ♦  

If Pn is the finest partition, the height h(P) of P is the length of the chain joining P and 

Pn. ♦  If Pk is a partition of k classes, h(P)=n-k. It can be shown that all chains joining P and 

P’ have the same number of elements, equal to the difference h(P) – h(P’) between their 

heights, and (i) if P and Q each cover R, then P ∨  Q covers both P and Q; (ii) h(P) + h(P’) ≥ 

h(P ∨  P’) + h(P ∧  P’). 



S. Levachkine & A Guzman-Arenas  10 of 31 

h1 

h2 h3 

h4 

a b c d e f 

2.3 Hierarchy, equivalent definitions 

Let E be a set of n elements, ℘ (E) the set of all subsets of E and P(E) the lattice of the 

partitions defined by the order relation P < Q. Let CH be a complete chain in the lattice, 

i.e. a chain linking the finest partition Pn, of n elements, to the coarsest partition P1=E.  

Now we can give two additional definitions of a hierarchy. 

Definition 3. A hierarchy is a set of partition classes constituting a complete chain, includ-
ing in particular the set E itself and the n subsets formed by the elements of E. ♦  

The passage from level k to level k-1 on CH corresponds to combining two classes. How-

ever, several levels can be passed over. Let P and Q be two non-consecutive partitions on 

CH, so that the classes of Q are either those of P or combinations of two or more classes of 

P. This leads to another direct definition. 

Definition 4. A hierarchy is a subset H of ℘ (E) such that (1) E∈ H, (2) if x and y are ele-
ments of E, then { x} ,{ y}∈ H, (3) if h and h’ are elements of H, then either h∩h’=∅  or 
h∩h’≠∅ , in which case either h⊂ h’ or h’⊂ h. ♦  Example: If E={ a,b,c,d,e,f} , then figure 
2 represents the hierarchy formed by the subsets { a} ,{ b} ,{ c} ,{ d} ,{ e} ,{ f}  with h1=E, 
h2={ a,b,c,d} , h3={ e,f}  and h4={ a,b,c}  

 

 

 

 

 

 

 

Fig. 2. Representation of a hierarchy. 

2.4 Ultrametrics and clustering 

A partial ordering of the elements of a hierarchy can be based on the inclusion relation 

and can be made a total ordering by the process of ascending a complete chain CH. In gen-

eral, the same hierarchy can be defined by several different chains; thus in the example of 



S. Levachkine & A Guzman-Arenas  11 of 31 

figure 2 we can use three chains CH1, CH2 and CH3, with their nodes numbered 0,1,2,3,4 

as follows: 

 
 CH1 a,b,c,d,e,f abc,d,e,f abc,d,ef abcd,ef  abcdef 
 CH2 a,b,c,d,e,f abc,d,e,f  abcd,e,f abcd,ef  abcdef 
 CH3 a,b,c,d,e,f a,b,c,d,ef abc,d,ef abcd,ef  abcdef 
        0        1       2      3      4 

 
Two elements of E occur in the same subset at a given node of CH, this being a parti-

tion of E. Given the chain, the node numbers characterize each pair of elements of E. We 

can now show how they can be used to define a special kind of distance. 

2 . 4 . 1  U l t r a m e t r i c  d i s t a n c e  f r o m  a  h i e r a r c h y  

 If i, j and k are three elements of a set E, the ultrametric distance δ is defined as a 

function of E×E in R+ as follows: 

•  δ(i,i)=0, 
•  δ(i,j)=δ(j,i), 
•  δ(i,j) ≤ max[δ(i,k),δ(j,k)] ♦                                      (1).  

 
It is easy to prove the following theorem: every triangle is either isosceles or equilat-

eral, with the base less than or equal to the equal sides. 

 
So we might define a distance between elements of E by means of a chain of partitions, and 

it will be clear that this is an ultrametric distance in the sense just defined. It will also be 

clear that infinity of ultrametric distances can be defined so as to be consistent with the 

order imposed by the chain CH, and we must remember that the same hierarchy can be 

specified by several different such chains. 

Conversely, an indexed hierarchy can be built (§2.4.2), given an ultrametric distance. 



S. Levachkine & A Guzman-Arenas  12 of 31 

2 . 4 . 2  C o n s t r u c t i n g  a  h i e r a r c h y  f r o m  a n  u l t r a m e t r i c  d i s t a n c e  

An ultrametric space is a couple consisting of a set E, finite or otherwise, and an ul-

trametric distance δ. In such a space, all triangles are isosceles or equilateral with the base 

less than or equal to the equal sides (see Theorem of §2.4.1). Conversely, if a generalized 

“distance” between any pair of elements of a set E is such that all triples give triangles hav-

ing this property, this distance has the ultrametric property given by relation (1). 

Let B be a sphere with center i and radius r, j an interior point and k a point on the sur-

face. By hypothesis δ(i,j)≤r, δ(i,k)=r. In the triangle (i,j,k) we must have δ(j,k)=r. This 

means that any internal point j is equidistant from all points on the surface, or that any in-

ternal point can be regarded as the center of the sphere. 

Similarly, it can be shown that if two spheres have a point in common then one must be 

included in the other; this follows from taking the common point as center for each sphere. 

Let h denote the set of points of a sphere; all these points are equidistant from each 

other. If j is a point not belonging to h then, by δ(i,j)≤δ(j,k), it is equidistant from all points 

of h and this constant distance is greater than r. Thus h is an element of a hierarchy H de-

duced from the ultrametric distance δ. We can use this result to define an algorithm for 

the construction of H from δ: 

Step 0. Set up the triangle table ∆ of ultrametric distances between the elements of E. 

Step 1. Find those elements hi,...,hk for which these distances are least (and equal, in the 

case of two elements only). Replace these in ∆ by hm. Recompute the distances between 

hm and the other elements hj: 

δ(hm,hm)=0,   δ(hm,hj)= δ(hi,hj) 

Step 2. If ∆ has more than one column, repeat step 1, else end. 



S. Levachkine & A Guzman-Arenas  13 of 31 

 

Example: Taking the example of figure 2, the successive tables (tables 1 to 5) are as fol-

lows. 

 
Table 1 a b c d e f 

a 0 1 1 3 4 4 
b  0 1 3 4 4 
c   0 3 4 4 
d    0 4 4 
e     0 2 
f      0 

 
Table 2 h4 d e f 

h4 0 3 4 4 
d  0 4 4 
e   0 2 
f    0 

 
Table 3 h4 d h3 

h4 0 3 4 
d  0 4 
h3   0 

 
Table 4 h4 h3 

h2 0 4 
h3  0 

 
Table 5 h1 

h1 0 
 

The hierarchy is the same as in Figure 2. 

A knowledge of ultrametric distances enables us to construct a hierarchy and therefore 

to form a sequence of partitions of decreasing fineness along a chain CH of the partition 

lattice, i.e. to perform a classification of variable fineness. What is now required is to find 

that ultrametric distance which meets the needs of the classification problem under con-



S. Levachkine & A Guzman-Arenas  14 of 31 

sideration. Since, in general, the data of a problem consist of distances in the ordinary 

sense, the requirement is to obtain an ultrametric distance from ordinary distance. 

2 . 4 . 3  O b t a i n i n g  a n  u l t r a m e t r i c  f r o m  a  m e t r i c  

Several algorithms that enable to us to derive an ultrametric which is close to the dis-

tance in terms of which the problem data have been proposed by Lance and Williams [14] 

and others [15]. 

2 . 4 . 4  D i s t a n c e s  a n d  l i n k a g e  e f f e c t  

In this section we define the chain distance value that can lead to the linkage effect.  

Let E be a set of n elements and d(i,k) the distance between two elements. We define a 

path cjk as a sequence of elements (j,…,q,…,k) and Cjk as the set of paths cjk, and  

δ(cjk) = max
q

d(q,q+1).  

Thus δ is the length of the longest link in the path and we say that this defines the length of 

the path. Using the minimax criterion, in the set Cjk we find a path of shortest length, δ(j,k) 

say. This is an ultrametric distance, which we call the chain distance value. ♦  

 The use of a chain distance can lead to linkage effect, which appears in the nearest 

neighbor method of clustering (single linkage). 

2 . 4 . 5  C l u s t e r i n g  m e t h o d s  

The concepts of hierarchy, ultrametrics and clustering are closely linked. 

It is natural to place different objects into the same cluster according to a criterion of 

neighborhood. As this proceeds, the partitioning becomes progressively less fine and new 

objects h of a hierarchy H, are created [11]. Use of the nearest neighbor or single linkage 

criterion can sometimes be dangerous because of the linkage effect, and for this reason we 



S. Levachkine & A Guzman-Arenas  15 of 31 

have to look for more satisfactory criteria which will take into account, among other things, 

the problem context under consideration. This is handled in next section. 

2.5 Conclusions 

Distances and ultradistances help us to establish a partial or total order among elements 

x,y,z,u ∈  E, this is written (x,y) ≤ (z,u) meaning that x resembles y more than z resembles 

u. If the number of elements of E is large, the establishment of this order may be difficult, 

so that a practical manner to order them (at least partially) is to define a numerical function 

of similarity or dissimilarity (confusion) that can be computed in terms of the attributes of 

every element of E: the dissimilarity (confusion) λ(x,y) will be smaller the more closely x 

resembles y. Moreover, this function may not be a distance. That is developed in next Sec-

tion. 

3. Properties and functions on hierarchies 

I ask for furniture, and they bring me a chair. Is there an error? Now, I ask for a saw, 

and a tool is brought. Can we measure this error? Can we classify or organize these values? 

Yes, by using hierarchies (figure 1). In this section, we measure the error (called confusion) 

when one symbolic value is used instead of another (the intended or correct value). 

3.1 Confusion in using r instead of s, for a hierarchy H 

If r, s ∈  H, then [12] the confusion in using r instead of s, written conf(r, s), is: 

•  conf (r, r) = conf (r, any ascendant_of (r)) = 0.  
•  conf (r, s) = 1 + conf (r, father_of(s)). ♦ 3  

                                                           
3 Without loss of generality we can define conf (r,s) = [1 + conf (r, father_of(s))] / height (H). In this case, 1 ≥ conf (r,s) ≥ 
0 



S. Levachkine & A Guzman-Arenas  16 of 31 

LIVE BEING 

ANIMAL PLANT 

CITRIC pine MAMMAL 

orange 

snake 

lemon dog cat 

bird 

To measure conf, count the descending links from r (the replacing value) to s (the replaced 

or intended value). conf is not a distance, nor ultradistance. Example: conf(r, s) for the hi-

erarchy H2 of figure 3 is given in Table 6. 

 

Fig. 3. A hierarchy H2 of living creatures. 

 

conf live 
being 

ani-
mal 

plant mam
mal 

bird snake citric pine cat dog lemon orange 

live 
being 

0 1 1 2 2 2 2 2 3 3 3 3 

ani-
mal 

0 0 1 1 1 1 2 2 2 2 3 3 

plant 0 1 0 2 2 2 1 1 3 3 2 2 
mam
mal 

0 0 1 0 1 1 2 2 1 1 3 3 

bird 0 0 1 1 0 1 2 2 2 2 3 3 
snake 0 0 1 1 1 0 2 2 2 2 3 3 
citric 0 1 0 2 2 2 0 1 3 3 1 1 
pine 0 1 0 2 2 2 1 0 3 3 2 2 
cat 0 0 1 0 1 1 2 2 0 1 3 3 
dog 0 0 1 0 1 1 2 2 1 0 3 3 
lemon 0 1 0 2 2 2 0 1 3 3 0 1 
orange 0 1 0 2 2 2 0 1 3 3 1 0 

 

Table 6. Confusion in using row r instead of column s for the live beings of H2. 

s 

r 



S. Levachkine & A Guzman-Arenas  17 of 31 

The confusion thus introduced catches the hierarchy semantics and resembles reality. 

For example, the error when using plant instead of live being is 0, since all plants are live 

beings. conf(plant, live_being) = 0: there is no error if they ask you for a live being and you 

give them a plant. Giving a live being when asked for a plant has error 1; conf (live_being, 

plant) = 1. The confusion among two brothers, such as orange and lemon, is 1. The confu-

sion when using the father instead of the son is 1. Using a son instead of the father gives 0. 

conf is not a symmetric function. Using specific things instead of general things produces 

low errors, see the column animal. Using general things instead of specific things produces 

high errors, see the row live_being. Using any descendant of a node n instead of n produces 

no error: observe the 0’s in column plant. The lower triangular half has smaller errors than 

the upper triangular half of the table 4. 

3 . 1 . 1  C o n f u s i o n  c o n f b  f o r  h i e r a r c h i e s  t h a t  a r e  f o r m e d  b y  b a g s  

Sometimes, the sizes of the sets that form the hierarchy matter. For example, for the hi-

erarchy soldier ∝  {male_soldier, female_soldier}, our definition of conf would yield 

conf(soldier, male_soldier) = conf(soldier, female_soldier) = 1, when these numbers should 

be something like conf(soldier, male_soldier) = 0.02, conf(soldier, female_soldier) = 0.98, 

since approximately 98% of soldiers are male. 

A percentage hierarchy H of E is a hierarchy in which the number of elements of E in 

each set of H is known. ♦  The nodes of H are not sets, but bags: unordered collection 

where repetitions are allowed. 

                                                           
4 A loss of context appears if we use an ultrametric distance or any other symmetric function: these triangular parts 
would be equal. 



S. Levachkine & A Guzman-Arenas  18 of 31 

For bags, the confusion in using r instead of s, confb(r, s) should take into account the rela-

tive popularity of s in r, that is, confb(r, s) = 1 – relative proportion of s in r. This is made 

precise in the following definition. 

 

For percentage hierarchies, confb(r, s) = 1 – (|E∩r∩s| / |E∩r|). ♦  It is one minus the number 

of elements of E that are present in r∩s, divided by the number of elements in E present 

in r. The intersection ∩ preserves repeated elements. 

Example: For E = NorthAmerica(330M) ∝  {Canada(30M) ∝  {French_Canada(5M), Eng-

lish_Canada(25M)}, USA(200M), Mexico(100M)}, where the population in millions is 

indicated, we show in table 7 the confusion confb. For instance, confb(Canada, NorthAme-

rica) = 1 – (30M/30M) = 0; confb(NorthAmerica, Canada) = 1-(30M/330M)=1-.09 = 0.91. 

See table 7. 

confb N C U M Fr_C En_C 
N 0 0.91 0.39 0.7 0.99 0.93 
C 0 0 1 1 0.63 0.17 
U 0 1 0 1 1 1 
M 0 1 1 0 1 1 
Fr_C 0 0 1 1 0 1 
En_C 0 0 1 1 1 0 

 

Table 7. The confusion confb(r, s) using row r instead of column s is shown for 
bags NorthAmerica, Canada, USA and Mexico, represented as N, C, U, and M. 

3 . 1 . 2  C o n f u s i o n  c o n f L  f o r  h i e r a r c h i e s  t h a t  c o n t a i n  l i s t s  

Sometimes, there is an order among the symbolic values that form a partition, as in H3 

={microscopic, tiny, small, medium, large, gigantic}. In this case, some nodes in the hier-

archy are lists (ordered sets), not just sets. For these, the confusion among two brothers 

should not be 1 (as the ordinary definition of conf will say), but a number between 0 and 1 



S. Levachkine & A Guzman-Arenas  19 of 31 

related to the proximity of the two brothers. For a hierarchy composed of sets, some of 

which have an ordering relation, the confusion in using r instead of s, confL(r, s), is defined 

as follows: 

•  confL(r, r) = confL(r, s) = 0, when s is any ascendant of r. 

•  If r and s are brothers,  

confL(r, s) = 1 if the father is not an ordered set; else, 

confL(r, s) = the relative distance from r to s = the number of steps needed to jump from 

r to s in the ordering, divided by the cardinality-1 of the father.  

•  confL(r, s) = 1 + confL(r, father_of(s)). ♦  Example: Refer to hierarchy H3. confL (micro-

scopic, tiny)=1/5; confL(microscopic, small)=2/5; confL(microscopic, gigantic)=1. 

The rest of the paper will derive results for conf; those for confb and confL can be simi-

larly derived. 

3.2 The set of values that are equal to another, up to a given confusion 

A value u is equal to value v, within a given confusion ε, written u =ε v, iff conf(u, v) ≤ 

ε. ♦  v is the “correct” or intended value. It means that value u can be used instead of v, 

within error ε [12]. Example: Refer to figure 3. The set of values equal to CITRIC with 

confusion 0 is {CITRIC orange lemon}. The set of values equal to CITRIC with confu-

sion 1 is {CITRIC orange lemon PLANT pine}. The set of values equal to cat with con-

fusion 2 is {ANIMAL MAMMAL bird snake dog cat}. Notice that =ε is neither sym-

metric nor transitive. 



S. Levachkine & A Guzman-Arenas  20 of 31 

3 . 2 . 1  P r e d i c a t e s  o n  v a l u e s  o f  h i e r a r c h i c a l  v a r i a b l e s  

An object can be characterized by several (variable, value) pairs, and some of the vari-

ables may be hierarchical, such as (Sue (lives_in Canada) (trades_in furniture) (size 

small)). It is thus appropriate to define predicates that may hold for such objects. 5 

In this section we extend the notion of predicate to “predicate that holds within a confu-

sion”, by defining the set S of objects that satisfy predicate P within a given confusion ε. 

P holds for object o with confusion ε, or P holds for o within ε, iff 

•  If P is formed by non-hierarchical variables, iff P is true for o. 

•  For pr a hierarchical variable and P of the form (pr c), iff for value v of property pr 

in object o, v =ε c (if the value v can be used instead of c with confusion ε).6 

•  If P is of the form P1 ∨  P2, iff P1 holds for o within ε or P2 holds for o within ε. 

•  If P is of the form P1 ∧  P2, iff P1 holds for o within ε and P2 holds for o within ε. 

•  If P is of the form ¬ P1, iff P1 does not hold for o within ε. ♦  

It is easy to see that if P1 holds for o within ε1 and P2 holds for o within ε2, then P1 ∨  P2 

holds for o within min(ε1, ε2), whereas P1 ∧  P2 holds for o within max(ε1, ε2). 

Example (refer to hierarchies H1 and H2 above):  

Let the predicates be  U = (trades_in furniture) ∨  (owns dog),  
V = (trades_in furniture) ∧  (owns dog), 
W = ¬  (trades_in apparel), 
X = ¬  (owns bird) 

 

                                                           
5
 For variables that are not hierarchical, a match in value means conf = 0; a mismatch means conf = ∞. 



S. Levachkine & A Guzman-Arenas  21 of 31 

and objects be  (Carl (trades_in furniture) (owns snake)), 
(Don  (trades_in apparel) (owns citric)), 
(Ed (trades_in shoe)  (owns ANIMAL)), 
(Fred (trades_in table) (owns orange)), 
(Gal  (trades_in tool) (owns PLANT)), 
(Hal (trades_in hammer) (owns dog)). 

 

Then we have the results of table 8. 

 U holds within ε 
for: 

V holds within ε
for: 

W holds within ε
for: 

X holds within ε 
for: 

ε = 0 Carl, Fred, Hal (nobody) Carl, Fred, Gal, 
Hal 

(all) 

ε = 1 (all) Hal (nobody) Don, Fred, Gal 
ε = 2 (all) Carl, Ed, Hal (nobody) (nobody) 

 

Table 8. How the predicates U, V, W, X hold for several objects. 

3 . 2 . 2  F u l f i l l m e n t  o f  a  p r e d i c a t e  b y  a n  o b j e c t  

In above example, how well an object such as Ed fulfils a predicate such as V? Looking 

at column V in table 8, V starts holding for o at ε=2. No smaller ε will do. 

Object o ε-fulfills predicate P at threshold ε, if ε is the smallest number for which P holds 

for o within ε. ♦  It is a non-negative integer defined between an object and a predicate. 

The closer is ε to 0, the “tighter” P holds for o. Compare with the membership function 

for fuzzy sets. 

3 . 2 . 3  C o n f u s i o n  b e t w e e n  t wo  o b j e c t s  

Similar to §3.1, for two objects (o (pr1 v1) (pr2 v2).. (prk vk)) and (o’ (pr1 v1’) (pr2 v2’)… 

(prl vl’)) with the same hierarchical variables, we can define the confusion when object o is 

                                                                                                                                                                                 
6
 c is the intended or correct value, and it is assumed to be a value (a constant) in the hierarchy of pr. If c is a constant 

that does not belong to the hierarchy of pr, then (pr c) is always false. If c were not a constant, but another variable, so 
that P is of the form (pr var), for instance (size  trades_in), [“I want the objects where size and trade_in  have the same 
value”] the value is false unless pr and var take values on the same hierarchy. In this case, the value of var is the 
intended value. 



S. Levachkine & A Guzman-Arenas  22 of 31 

used instead of o’ as the sum of the individual confusions conf(vi, vi’) between their respec-

tive qualitative values. o’ is the intended or correct object. 

Confusion when object o is used instead of o’. CONF (o, o´) = Σi conf(vi, vi’). ♦  CONF 

(written in capital letters) is not symmetric, and is not a distance. 

3.3 Confusion between variables (not values) that form a hierarchy 

What could be the error in “Carl is a relative of Mary,” if all we know is “Carl knows 

Mary”? And if what we know is “Carl is the grandfather of Mary”? That is, in addition to 

qualitative values forming a hierarchy, it is possible for the variables themselves to form a 

hierarchy. Example: hierarchy H4. 

handles ∝   
{trades_in  ∝   {buys sells rents} 

     carries      ∝  {transports_by_surface flies ships} 
    maintains ∝  {oils fixes } 

} 
 

When considering H4 for (Carl (trades_in furniture) (owns snake)) of example § 3.2.1, 

he surely (handles furniture) with confusion 0, (rents furniture) with confusion 1, (rents 

merchandise) with confusion 1, and (rents table) with confusion 1+1=2. Thus, we can ex-

tend the definition of predicate with confusion to have variables that are members of a hier-

archy, by adding another bullet to the definition of §3.2.1, thus: 

•  If P is of the form (var c), for var a variable member of a hierarchy, iff ∃  variable var2 

for which (var2 c) holds for o within ε – conf (var2, var), where var2 also belongs to the 

hierarchy of var.♦  The confusion of the variables adds to the confusion of the values. 



S. Levachkine & A Guzman-Arenas  23 of 31 

Example: For (Don (trades_in apparel) (owns citric)) of E2, predicate (handles apparel) 

holds with confusion 0; (trades_in apparel) holds with conf=0, (buys apparel)  holds with 

conf=1, (buys shoe) holds with conf= 1+1=2, and (buys shoe) ∧  (owns lemon) with confu-

sion = max(2, 1) = 2. Predicate (transports apparel) holds for Don with conf=1+1=2. 

3.4 Confusion confdo for values in different ontologies 

It is possible that values v1 and v2 belong to hierarchies coming out of different element 

sets E1 and E2 expressed as hierarchies H1 and H2. If these hierarchies can be converted into 

ontologies (to be called O1 and O2), and thus map values v1 and v2 (v2 is the intended 

value) into objects (concepts) c1∈ O1 and c2∈ O2, we can use function sim of [13] to find 

confdo(v1, v2). Function sim(c1, O1, O2) finds sv, a number between 0 and 1 expressing the 

similarity between a concept c1 in O1 and its most similar  concept c’2∈ O2. It also finds 

such c’2. Thus, to find confdo(v1, v2), execute these steps: 

1. Find c’2=sim(c1, O1, O2), the concept c’2 most similar to c1, as well as sv. 

2. Since c2 and c’2 belong to the same ontology O2, find l = length of the path going 

from c2 to c’2 in the OB tree. 

3. confdo(v1, v2) = (1+l)/sv.♦  The steps are depicted in figure 4. See also [21]. 



S. Levachkine & A Guzman-Arenas  24 of 31 

v1 
c1 v2 

c2

c'2 
O1 (0) 

(0)

(1)

l
(2)

O2

 

 

 

 

 

 

Figure 4. Steps to find confdo(v1, v2) for values in different ontologies: (0) map 
v1, v2 into c1, c2; (1) find c’2 and sv; (2) find l; (3) (not shown) compute (1+l)/sv. 

4. Discussion 

Starting the discussion, let us look again at the ordering and similarity of the elements 

of finite sets. A reflexive and transitive relation can be defined for the set F of all pairs of 

elements in E x E, which is called a partial order. With x,y,z,u ∈  E, this is written (x,y) ≤ 

(z,u) meaning that x resembles y more than z resembles u. Such a relation does not neces-

sarily apply over the complete set F because they may be pairs of elements that are really 

not comparable. If it does apply over the complete set it becomes a total order.  

It is difficult in practice to set up such a partial order if the number of elements of E is 

large, and if it is possible it is difficult to make this order without running the risk of gener-

ating contradictions. In fact, the only practical way to establish a partial order is to define 

a numerical function of similarity or dissimilarity (confusion) that can be computed in 

terms of the attributes of every element of E: the dissimilarity (confusion) λ(x,y) will be 

smaller the more closely x resembles y. 

The same partial order can be generated by any of an unlimited number such functions. 

Some dissimilarity functions, however, may not be distances (Cf. §3). Thus, by definition, a 



S. Levachkine & A Guzman-Arenas  25 of 31 

distance d(x,y) satisfies: (a) d(x,x) = 0; (b) d(x,y) = d(y,x); (c) d(x,y) ≤ d(y,z) + d(x,z). Re-

lation (a) can be satisfied by making min λ = 0, which is simply a change of origin, and for 

(b) it is sometimes necessary to make λ symmetrical. Relation (c) may not hold, however, 

although it can be made to hold by adding a sufficiently large constant λ0 to the values, 

whilst retaining λ(x,x) = 0. The corresponding partial ordering will not be changed. Thus 

the following general statement can be made: it is always possible for a finite set E to make 

simple modifications that will transform a dissimilarity into a distance measure without 

affecting the partial order. 

In early approaches [25], the measures of similarity between variables have been con-

sidered using Kendall, Hamming, Russell & Rao, Jaccard, Kuzlinsky, Yule, and other dis-

tances, and a contingency table. These result, however, insufficient for variables that take 

symbolic values. 

In context of the identification problem, distance measures for objects and concepts 

have been proposed too. The idea of classification carries with it in the implication that a 

descriptor or symbolic description can be defined for each class and will in some sense 

be representative of the class; possible descriptors are a symbolic description of the class; a 

feature represented as a relational structure. If X is the representation of an object, we can 

define a concept Ai as an entity which is such that an ordering of the couples (X,Ai) can be 

established for i = 1,...,k. A concept is not necessarily associated with one particular class; 

this may be so, e.g. in the case of descriptors, but in other cases the concepts may overlap, 

as for example for probabilities. The “distance” D(X,Ai) between an object and a concept 

can be used effectively as a measure of similarity or dissimilarity. In other words, the “ob-

ject-concept” distance D is used as a characteristic function. The following functions have 



S. Levachkine & A Guzman-Arenas  26 of 31 

been used: probability, fuzzy assignment, inertia and potential, nearest neighbors (q-NN), 

etc. The main problem with these approaches is that they often (always) omit the context of 

a classification problem under consideration. 

Let f(X;A) be a measure of similarity or dissimilarity between the representation of an 

object X and a concept A. Such a measure limited to a single A would be of little interest; if 

for example, A were the symbolic description of a class, there would clearly be at least two 

classes, A and ¬ A (“not-A”). However, it can happen that the concepts A do not coincide 

with the classes. Let {Ai} be the set of concepts under consideration: the assumption that 

these concepts can be grouped together into a set implies that the set operations of union 

(∪ ) and intersection (∩) can be applied. The Ai are not necessarily disjoint, i.e. several can 

apply simultaneously to a single object, although in the very special case that each identi-

fies one class they are clearly disjoint. Let @ be the algebra generated by the Ai and opera-

tions: ∩, ∪ , and ¬ ; the elements of @ form the interpretation space. If the concepts Ai are 

expressed by predicates, the operations are written as ∧  and ∨  respectively and we have 

Stone’s theorem: all distributive logic systems are homomorphic to a distributive lattice of 

subsets of a set. We may recall that if p,q and r are predicates then, ∧  is distributive with 

respect to ∨ . 

Given f(X;A) and f(X;B), the fundamental question is: what are the values of f(X;A∩B) 

and f(X;A∪ B)? This can be answered in terms of two laws or operations: (i) an additive 

law ⊕ , homomorphic with ∪ , and (ii) a multiplicative law ⊗ , homomorphic with ∩ and 

distributive over ⊕ , and the answer is [15, 25]  (Cf. §3.2.1) 

f(X; A∪ B) = f(X;A) ⊕  f(X;B), 



S. Levachkine & A Guzman-Arenas  27 of 31 

f(X;A∩B) = f(X;A) ⊗  f(X;B)                                    (2). 

 
Union and intersection of ontologies and corresponding measures for them will be con-

sidered in a forthcoming work. 

4.1 Summing up 

We can emphasize in this summary the following key points: 

•  The context of a problem under consideration may be lost when attempting to define a 

distance on hierarchies of symbolic values (to measure closeness between hierarchical 

elements) that holds its partial (total) order. 

•  We show (§3.1) a way that takes into account the problem context, represents the in-

trinsic data semantics and expresses the similarity (dissimilarity) function in terms of 

the data attributes. 

•  Such approach permits to know the set of values that are equal to another up to a given 

confusion (§3.2), how close an object fulfils a predicate, and how close an object is to 

another (§3.2.3). 

•  Confusions and similarity functions for values in different ontologies can be defined as 

well (§3.4). 

4.2 Applications 

A) To queries, either to retrieve objects that hold for a predicate to a given threshold [5], 

or to measure the closeness of an object to a certain predicate (definition of §3.2.2). 

B) To handle partial knowledge. Even if we only know that Carl trades in furniture, we 

can productively use this value in precise searches (example of §3.2.1). 



S. Levachkine & A Guzman-Arenas  28 of 31 

C) As an approximation to the manner in which people use gradation of symbolic values 

(ordered sets), to provide less than crisp, but useful, answers. 

D) As a measure of the confusion between two values (§3.1), such as in the answer “the 

capital of Germany is Frankfurt” (Cf. Introduction), versus “the capital of Germany is 

sausage.” 

E) To compare attributes (as opposed to values) that are similar, but not equal, such as 

my_neighbor and my_acquaintance (Cf. §3.3). 

F) As an alternative to fuzzy sets, using Pε(o) (§3.2.2) as the membership function of a 

set. 

G) As a supervised pattern classifier, by using CONF(o, o’) of §3.2.3 between two ob-

jects. See also [18] for another approach to classifiers of qualitative data. 

H) As an alternative to data mining, where approximate answers are usually useful. 

I) As an accelerator of applications (G) and (H), if we store in cache memory the queries 

and results of previous tasks, and compare the (predicates of the) new queries to these 

cached data, before embarking in new searches [17]. 

5. Conclusions and suggestions for further work 

5.1 Conclusions 

The notions of hierarchy and hierarchical variable make possible to measure the confu-

sion when a value is used instead of another. This creates a natural generalization for predi-

cates and queries. The notions were introduced and developed for hierarchies formed by 

sets, but they can be extended to bags and lists, too. 



S. Levachkine & A Guzman-Arenas  29 of 31 

The concepts and examples given here have practical applications, since they mimic the 

manner in which people process symbolic values. 

In a subsequent paper we will describe a mathematical apparatus and further properties 

of functions defined in §3. See also [13, 22] for ontologies instead of hierarchies. 

5.2 Suggestions for further work 

1. Extend definition 3.2 (the sets of values equal to another value) to fuzzy sets. 

2. Extend definition 3.3 (confusion between variables, not values) to fuzzy sets. 

3. Extend queries of 3.2.2 to queries (search) in text, with the help of Clasitex [9]. 

4. Idem to queries in maps. 

5. Idem to queries in semi-structured data. 

Acknowledgments 

The advice of the referees was very useful. Helpful discussions were held with Prof. 

Victor Alexandrov, SPIIRAS-Russia, Dr. Jesus Olivares and Prof. Gilberto Martinez, CIC-

IPN. Work herein described was partially supported by NSF-CONACYT Grant 32973-A 

and Project CGPI-IPN 20010778. The authors have a SNI National Scientist Award. 

References 

1. V. Alexandrov. Developed systems in science, technique, society and culture. (Saint 
Petersburg State Technical University, Russia, 2000). In Russian. 

2. N.T. Bhin, A.M. Tjoa, and R. Wagner. Conceptual Multidimensional data model based 
on meta-cube. Lecture Notes in Computer Science 1909 (Springer-Verlag, 2000) 24-31 

3. A. Budanitsky and G. Hirst. Semantic Distance in WordNet: An Experimental, 
Application-oriented Evaluation of Five Measures. In: Proc. North American Chapter 
of the Association for Computational Linguistics (NAACL-2000), Pittsburgh, PA. 

4. V.P, de Gyves and A. Guzman-Arenas. A distributed digital text accessing and 
acquisition system. BiblioDigital.  SoftwarePro International. In: Lecture Notes in 
Computer Science 3061 (Springer-Verlag 2004) 274-283. 



S. Levachkine & A Guzman-Arenas  30 of 31 

5. V.P. de Gyves. Precision-controlled retrieval of objects with symbolic values, in a 
relational database. B. Sc. Thesis in preparation, UPIICSA-Instituto Politecnico 
Nacional. In Spanish. 

6. J. Everett, D Bobrow, et al. Making ontologies work for resolving redundancies across 
documents. Comm. ACM 45 (2) (2002) 55-60. 

7. D. B. Lenat and R. V. Guha. Building Large Knowledge-Based Systems. (Addison-
Wesley 1989). 

8. A. Gelbukh, G. Sidorov, and A. Guzman-Arenas. Document comparison with a 
weighted topic hierarchy, In: Proc. DEXA-99, 10th International Conference on 
Database and Expert System applications, Workshop on Document Analysis and 
Understanding for Document Databases (Florence, 1999) 566-570. 

9. A. Guzman. (1998) Finding the main themes in a Spanish document. Journal Expert 
Systems with Applications 14 (1/2) (1998) 139-148. 

10. A. Guzman, J. Olivares, A. Demetrio and C. Dominguez. Interaction of Purposeful 
Agents that use Different Ontologies. In: Lecture Notes in Artificial Intelligence 1793 
(Springer-Verlag 2000) 557-573. 

11. A. Guzman and S. Levachkine. Graduated errors in approximate queries using hie-
rarchies and ordered sets. In: Lecture Notes in Artificial Intelligence 2972 (Springer-
Verlag 2004) 139-148. 

12. S. Levachkine and A. Guzman-Arenas. Hierarchies Measuring Qualitative Variables. 
In: Lecture Notes in Computer Science 2945 (Springer-Verlag 2004) 262-274. 

13. A. Guzman and J. Olivares. Finding the Most Similar Concepts in two Different Onto-
logies. Lecture Notes in Artificial Intelligence 2972 (Springer-Verlag 2004) 129-138. 

14. G.N. Lance and W.T. Williams. Mixed-Data Classificatory Programs I – 
Agglomerative Systems. Australian Computer Journal 1 (1) (1967) 15-20. 

15. S. Levachkine. Qualitative data conversion and representation. MIEM (Moscow 
Institute of Electronics and Mathematics, 1994) Report (in Russian). 

16. D. Lin. An Information-Theoretic Definition of Similarity. Proc. 15th Int. Conf. on 
Machine Learning (ICML 1998) Madison, Wisconsin. 

17. G. Martinez-Luna. Incremental data mining. Ph. D. Thesis in preparation. Centro de 
Investigacion en Computacion, Instituto Politecnico Nacional. In Spanish. 

18. F. Martinez-T and A. Guzman. The logical combinatorial approach to Pattern 
Recognition, an overview through selected works. Pattern Recognition 34 (2001) 741-
751. 

19. M. Montes-y-Gómez, A. Lopez-Lopez, and A. Gelbukh. Information Retrieval with 
Conceptual Graph Matching. In: Lecture Notes in Computer Science 1873 (Springer-
Verlag 2000) 312-321. 

20. N. Noy, R.W. Fergerson and M.A. Musen. The knowledge model of Protégè-2000: 
combining interoperability and flexibility. Stanford Medical Informatics Technical 
Report, Stanford University, 2000 

21. J. Olivares. A model of interaction among purposeful agents with mixed ontologies and 
unexpected events. Ph. D. Thesis. Centro de Investigacion en Computacion, Instituto 
Politecnico Nacional, 2002. In Spanish. Available on line at 
http://www.jesusolivares.com/interaction/publica 



S. Levachkine & A Guzman-Arenas  31 of 31 

22. J. Olivares and A. Guzman-Arenas. Concept similarity measures the understanding 
between two agents. In Proceedings of NLDB 04. (2004) 

23. P. Resnik. Disambiguating Noun Groupings with respect to WordNet Senses. In: 
Armstrong, S. et al. (eds.), Natural Language Processing Using Very Large Corpora. 
(Kluwer Academic Publishing, Dordrecht, 1995) 77-98. 

24. P. Resnik. (1999) Semantic Similarity in a Taxonomy: An Information-Based Measure 
and its Application to Problems of Ambiguity in Natural Language. Journal of 
Artificial Intelligence Research 11 (1999) 95-130. 

25. J-C Simon. Patterns and operators. The foundations of data representation. (McGraw-
Hill, 1984 

26. WordNet: A Lexical Database for the English Language http:// 
www.cogsci.princeton.edu/~wn/