-
Text mining is a rapidly growing application of knowledge discovery in data (cf.
[
2
]). A special kind of data in the textual form is constituted by advertisements,
for example, real-estate ones. In case of advertisements, data have the form of
loosely coupled words (terms, concepts) rather than full, grammatically correct
sentences. Moreover, underlying data are of qualitative character. We can
distinguish several challenges posed by textual data: understanding data semantics
and semantic relations between them, considering the external knowledge in
processes of data classi cation, encoding textual data for classi ers working with
numerical data.
The semantic relations between concepts play an important role, among
others, in cognitive psychology, linguistics and currently also in computer science. In
[
4
], two general types of relations between words (concepts) were distinguished:
Paradigmatic relations (relations between words belonging to the same
grammatical category) and Syntagmatic relations (relations between words that go
together in a syntactic structure). Paradigmatically related words are, to some
degree, grammatically substitutable for each other. In our research, we are
interested in paradigmatic relations. As it will be shown later, in real-estate listings,
we do not analyze a semantic structure of sentences, but we try to derive some
knowledge about concepts (terms) included in them, for example, whether they
are synonyms, whether one concept can be replaced with another, for example,
more general one, etc.
In our research, we use the following taxonomy of types of semantic
relations (which is modeled on the project called Wikisaurus [
1
] aiming at creating
a thesaurus of semantically related terms): synonymy, antonymy, hyponymy/
hyperonymy (subclass - superclass), and meronymy/ holonymy (part - whole).
In the approach presented in this paper, we are interested in synonymy and
hyponymy/ hyperonymy. We will use the following notation: isSyn denotes
synonymy, (u; v) 2 isSyn means that "u is a synonym of v", isGen denotes
hyponymy, (u; v) 2 isGen means that "u is a hyponym of v" ("u is generalized
by v"), and isSpec denotes hyperonymy, (u; v) 2 isSpec means that "u is a
hyperonym of v" ("u is specialized by v"). Additionally, we take into
consideration a semantic relation called "being an instance". Being an instance concerns
an example (instance) of a given concept. This kind of relations is important in
mining real-estate listings because they include, for example, instances of places.
The knowledge about semantic relations between concepts is included in
ontologies. Our approach is based on the de nitions of ontology given by Neches et
al. [
5
] and Kohler [
3
]. That is, ontology is constructed on the basis of a controlled
vocabulary and the relationships of the concepts in the controlled vocabulary.
Formally, the ontology can be represented by means of graph structures. Let
O be a given ontology. An ontological graph is a quadruple OG = (C; E; R; ),
where C is a nonempty, nite set of nodes representing concepts in the ontology
O, E C C is a nite set of edges representing relations between concepts
from C, R is a family of semantic descriptions of types of relations (represented
by edges) between concepts, and : E ! R is a function assigning a semantic
description of the relation to each edge. Sometimes, we can also consider
subgraphs of ontological graphs, called by us local ontological graphs. In Figure 1,
exemplary ontological graphs representing the real-estate domain are shown.
a)
b)
In [
12
], information (decision) systems were proposed as the knowledge
representation systems. In simple case, they consist of vectors of numbers or symbols
(attribute values) describing objects from a given universe of discourse. In our
research, we are interested in mining textual data in the form of concepts (words,
terms). Therefore, in [
7
] and [
10
], ontologies were incorporated into information
(decision) systems, i.e., attribute values were considered in the ontological
(semantic) space. Information (decision) systems over ontological graphs can be
created in di erent ways. In [
7
], two approaches were mentioned. In the rst
one, attribute values are concepts from ontologies assigned to attributes - a
simple information (decision) system over ontological graphs. In the second one,
attribute values are local ontological graphs of ontologies assigned to attributes
- a complex information (decision) system over ontological graphs. In case of
real-estate listings, whether a client was interested in a given property can be
considered as a decision attribute.
The main goal of this paper is to show how to transform real-estate listings
into simple information (decision) systems over ontological graphs. This is a very
important preprocessing step that can be generally depicted as it is shown in
Figure 2. We can distinguish three main steps:
{ Stemming - de ning basic grammatical forms (roots) for particular words
existing in advertisements, for example, using a quite popular Porter stemming
algorithm [
13
].
{ Attributation - assigning concepts (built from words) existing in
advertisements to proper attributes as their values, according to de ned ontological
graphs.
{ Deinstantiation - replacing instances existing in advertisements with the
most speci c concepts (with respect to the hyponymy / hyperonymy
relation) whose instances they are.
Deinstantiation is an important step if we are interested in a more general
knowledge derived from real-estate listings, for example, some client is interested only
in houses in a village (not a particular one).
Let us consider, as an example, the following advertisement: "For sale,
Warsaw, Poland, Villa, bedrooms: 4, 437 m2". Let us also assume that we take into
consideration three attributes: Transaction, Place, and Property with proper
ontological graphs assigned to them. After performing our procedure, the
advertisement becomes one object (row) in an information (decision) system:
U=A T ransaction
P lace
P roperty
u1
Having an information (decision) system over ontological graphs, we can
apply di erent machine learning and data mining methods to extract some valuable
knowledge. In our previous papers, for such systems, we considered: rough sets
[
9
], decision rules [
8
], decision rules based on the DRSA approach [
6
], neural
networks [
11
].