With the Semantic Web progress many independently developed distinct domain ontologies have to be shared and reused by a variety of applications. The use of ontologies in information retrieval applications allows the retrieval of semantically related documents to an initial users' query. This work presents a fuzzy information retrieval model for improving the document retrieval process considering a knowledge base composed of multiple domain ontologies that are fuzzy related. Each ontology can be represented independently as well as their relationships. This knowledge organization is used in a novel method to expand the user initial query and to index the documents in the collection. Experimental results show that the proposed model presents better overall performance when compared with another fuzzy-based approach for information retrieval.
With the grown availability of information many research has been made in order
to provide intelligent ways to easy the information access. One point is to treat
not only the lexical information features but also to consider its semantics, that
is, the meaning attached to it. Within this approach the usage of ontologies to
organize the knowledge and to express semantic meaning has gaining attention.
An information retrieval system stores and index documents such that when
users express their information need in a query the system retrieves the related
documents associating a score to each one. The higher the score the greater the
document relevance [
When working with specific domain knowledge this problem can be overcome
by incorporating a knowledge base which depicts the relationships between index
terms into the existing information retrieval systems. Knowledge bases can be
manually developed by domain experts or automatically constructed from the
knowledge in the document collection [
We present an information retrieval model which is supported by fuzzy
related lightweight ontologies each one representing a distinct knowledge domain.
The model provides means to represent each ontology independently as well as
their relationships. This way existent ontologies can be reused in the model.
Based on the knowledge from the ontologies the system carries an automatic
fuzzy query expansion. The documents are indexed by the concepts in the
ontologies allowing the retrieval by their meaning. The documents do not need to be
indexed by each ontology concepts as in a faceted approach. Given a query with
concepts from an initial domain new semantically related documents indexed
by other domain ontologies can be retrieved based on the ontologies
relationships. The results obtained with the proposed information retrieval model are
compared with the results obtained using just the user’s entered keywords and
with the results obtained by another fuzzy information retrieval system [
Some information retrieval models that encode knowledge among terms in
order to improve performance are presented. In context sensitive semantic query
expansion [
The multi-relationship fuzzy concept network information retrieval model [
The fuzzy ontological relational model [
An ontology is a concept set Dk = {ck1, ck2, · · · cky} where 1 ≤ k ≤ K, K is
the domains number and y = kDkk is the concepts number in each domain. The
concepts inside an ontology are organized as a taxonomy and are related by fuzzy
specialization association (S) and fuzzy generalization association (G). The fuzzy
generalization association is the inverse of the fuzzy specialization association.
A concept is regarded as a fuzzy generalization of another concept if it consists
of that concept or it includes that concept in a partitive sense. A concept is
regarded as a fuzzy specialization of another concept if it is part of that concept
or it is a kind of that concept. Concepts pertaining to distinct ontologies are
related by fuzzy positive association (P). The fuzzy positive association denotes
concepts related by a spatial, causal or similarity relation in some contexts.
Definition 1. Consider two distinct concept domains sets Di and Dj .
1. Fuzzy positive association is a fuzzy relation: (RPij : Di × Dj → [
The implicit relationships between concepts from the same domain are given by the transitive closure of the fuzzy generalization and fuzzy specialization associations. The transitive closure of the associations RiG and RiS where 1 ≤ i ≤ K, results the relations RGi and R∗Si respectively.
∗
Definition 2. The transitive closure R∗ of a fuzzy relation R can be determined
by an iterative algorithm that consists of the following steps:
1. Compute R´= R ∪ [wet(R ◦ R)] where wet ∈ [
The (R◦R) means the composition between two fuzzy relations. The composition
between two fuzzy relations [
R(x, z) = (P ◦ Q)(x, z) = my∈aYx min [P(x, y), Q(y, z)] . (1)
The weight wet, with empirical values 0 < wet < 1, penalizes the association strength between distant concepts in the ontology. As the distance between concepts increase their association values decrease. Concepts with higher strength value are considered to have stronger meaning association. In order to discard concepts associations with lower strength value a boundary b establishes the minimum value such that the corresponding association is to be considered.
The documents dl are represented by the DOC set where 1 ≤ l ≤ kDOCk. A
fuzzy relation Uj (dl, cjy) = uly ∈ [
A query is expressed with concepts from distinct domains connected by logical operators. The query is transformed into the conjunctive normal form and is represented by sub-queries connected by the AND logical operator. Each subquery is composed by a set of concepts connected by the OR logical operator. Given the domains D1 = {c11, c12, c13} and D2 = {c21, c22, c23, c24} a valid query in this format would be q = (c11 ∨ c22) ∧ (c13 ∨ c24). Once the query is in the conjunctive normal form each sub-query is performed independently and retrieves a document set. The intersection of the document sets is the final result of the query. Therefore, in the sequence, only aspects related to the sub-query are presented. The documents are associated to the domain concepts using distinct relations. To consider this the sub-queries are partitioned to take the concepts from each domain separately. Each partition is a set with dimension equal to the associated domain concepts number and is composed by values that indicates the presence (1) or absence (0) of the concept in the query. A sub-query q is partitioned in qi sets where 1 ≤ i ≤ K and K is the domains number. In the previous example the sub-query q = (c11 ∨ c22) is partitioned as q1 = [1 0 0] and q2 = [0 1 0 0]. 3.2
Query expansion is performed in two phases. In the first one each partition qi, from the initial sub-query q, is expanded to consider the relations between the domain Di associated to the partition and the other domains from the knowledge base. For each partition qi new K − 1 sets are generated each one containing concepts from the others domains Dj , j 6= i, 1 ≤ i, j ≤ K associated to concepts present in qi. This process generates a new expanded query denoted qent. The first expansion is translated in (2). The variable i refers to the domain of the partition qi and the variable j refers to the remaining domains from the knowledge base.
K K qent = [ [ i=1 j=1 qi P j = i wP qi ◦ Rij j 6= i .
(2)
To expand the query to consider other domains the fuzzy positive association
RiPj between concepts from the domains Di and Dj is used. The model allows
to associate a weight wP ∈ [
qentiTj
K K qexpT = [ [ max wrwrRrRj r∗◦j q◦eqnetniTjtiTj j = i .
i=1 j=1 ∗ j 6= i
Considering the knowledge inside the domains each transposed partition
qentiTj, 1 ≤ i, j ≤ K is expanded to take into account the fuzzy generalization
and fuzzy specialization associations between the concepts from their domain
Dj. The model allows to associate a value wr ∈ [
The documents relevance is given by the similarity function between the documents representation and the expanded fuzzy sub-query qexpT. The similarity is calculated by the product between the relations Uj with each partition qexpiTj, as in (4), resulting the retrieved documents set V.
K K V = [ [ i=1 j=1
Uj × qexpiTj .
(3) (4)
Each relation Uj associates the collection documents to the Dj domain
concepts, where 1 ≤ j ≤ K. The set qexpiTj represents the expansion of the concepts
from the partition qi to the Dj domain where 1 ≤ i, j ≤ K. It is constituted from
the Dj domain concepts and its values indicates the degree the concepts from
domain Dj are associated to the concepts in partition qi. The arithmetic
product Uj × qexpiTj indicates the documents associated to the Dj domain that are
related to the qi partition. The S symbol designates union and denotes the max
operator. The arithmetic product adjusts the associations of the documents to
the concepts (expressed in the relations Uj) by the strength of the relationships
between concepts present in qexpiTj. The V (vl) set represents all the documents
in the collection and its value vl ∈ [
The model evaluation uses a document collection sample referring to the agrometeorology domain in Brazil, a query set, a lightweight ontology referring to the geographical brazilian territory and a lightweight ontology referring to the climate distribution over the brazilian territory. Both ontologies are manually constructed. 4.1
To construct the ontologies the brazilian map from Fig. 1 is considered. For both ontologies the fuzzy generalization association and the fuzzy specialization association relates the spatial relationship between the entities they refer to. As ontologies are considered as crisp ones the fuzzy generalization and the fuzzy specialization relationships assume values in the set {0, 1} denoting the existence (1) or absence (0) of the relationship between concepts. The positive relationships between distinct ontologies concepts are calculated as fuzzy ones.
The first ontology refers to the brazilian territory, say domain D1, with three levels. The root node is labeled ’Brazil’, the descendant nodes are labeled with brazilian regions and each region node has the respective brazilian state nodes as descendants. For the brazilian territory ontology the North Region is part of Brazil country so the fuzzy specialization association value is
S R1 (Brazil, North Region) = 1.0. This means that Brazil concept is specialized by North Region concept. As the fuzzy generalization association is the inverse of G the fuzzy specialization association then R1 (North Region, Brazil) = 1.0. This means that North Region concept is generalized by Brazil concept. Figure 2 shows a sample of the brazilian territory ontology.
The second ontology refers to the climate distribution over the brazilian territory, say domain D2. The root node is labeled ’Climate’, the root descendant nodes are labeled with brazilian zonal climates and each zonal climate has the respective associated Ko¨ppen climate nodes as descendants. Figure 2 shows a sample of the brazilian climate ontology.
The relationship between ontologies is established by the distribution of climate over brazilian territory as observed in the map. The relationship is settled in two levels. The first one is between brazilian regions and zonal climates and the second one is between brazilian states and Ko¨ppen climates. The dashed lines in Fig. 2 illustrate both relationships levels. The first level of fuzzy positive relationship is between brazilian regions and zonal climates. The value of relationships is given by mapping scanning. For example, the tropical zonal climate occurs in North Region. The amount of tropical climate in Brazil is 59, 811 pixels. The amount of tropical climate in North Region is 30, 616 pixels. So the fuzzy positive association between North Region and tropical climate is given by relation value R1P2 (North Region, Tropical) = 0.51. This means that North Region concept implies Tropical climate concept by fuzzy positive association with strength 0.51. On the other hand the North Region extent is 43, 737 pixels so the fuzzy positive association between tropical climate and North Region is R2P1 (Tropical, North Region) = 0.70. This means that Tropical concept implies North region concept by fuzzy positive association with strength value 0.70.
The second level of fuzzy positive relationship is between brazilian states and Ko¨ppen climates. The total amount of Cfb Ko¨ppen climate in Brazil is 1, 781 pixels. The amount of Cfb Ko¨ppen climate in Santa Catarina state is 693 pixels. So the fuzzy positive association between Santa Catarina State and Cfb Ko¨ppen climate is R1P2 (Santa Catarina, Cfb) = 0.39. This means that Santa Catarina concept implies Cfb Ko¨ppen climate concept by fuzzy positive association with strength value equal to 0.39. On the other hand the Santa Catarina state extent is 900 pixels so the fuzzy positive association between Cfb Ko¨ppen climate and Santa Catarina state is R2P1 (Cfb, Santa Catarina) = 0.77. This means that Cfb Ko¨ppen climate concept implies Santa Catarina state concept by fuzzy positive association with strength value equal to 0.77. In the information retrieval process if a user constructs a query composed with Cfb concept the query expansion process, based on the R2P1 relation, points that Santa Catarina concept is related with Cfb concept with strength 0.77 and this concept is added to query. Considering that the document collection is about agrometeorology domain this indicates that documents indexed with Santa Catarina concept can be possibly a relevant answer, even if the documents are not indexed with the Cfb concept itself. As Cfb ko¨ppen climate covers a fraction of 0.77 from Santa Catarina state then documents related to Santa Catarina concept can contain aspects related to Cfb Ko¨ppen climate even this concept itself is not present in the documents.
The Alignment Format Level 0 [
P positive association R12 between the brazilian territory ontology, D1, and the brazilian climate ontology, D2, is presented in the following. The proposed model performance is compared with a similar approach, that is, the multi-relationship fuzzy concept network information retrieval model presented in Sect. 2. The experiment also tests the use of the query expansion method in the Apache Lucene text search engine. The Apache Lucene allows boosting a search concept increasing the relevance of documents indexed by the concept.
The document collection is composed of a sample of 128 documents selected
from a base of 17,780 documents from the agrometeorology domain. This sample
considers documents containing just one of the concepts from the ontologies as
well as a combination of concepts from both ontologies. The queries set contains
35 queries considering just one concept from each ontology or two concepts from
both ontologies connected with AND or OR boolean operators. For each query
the relevant documents from the sample document collection are selected by a
domain expert. Several experiments were ran considering many combinations
of the weights wet, t ∈ {S, G} and wr, r ∈ {S, G, P}. After many tests all the
models showed a behavior tendency concerning the precision and recall measures.
Recall is the fraction of the relevant documents which are retrieved over all
relevant documents in collection related to query q and precision is the fraction
of the retrieved documents which are relevant to a query q over all documents
in the answer set [
As the addition of general concepts tends to add more noise in the search results then a lower weight value is assigned to fuzzy generalization association like wG = 0.3. A higher value is assigned to fuzzy specialization association like wS = 0.7. Following the same reasoning for transitive closure calculation, the tests showed that best results are achieved with lower values assigned to weight weG and higher ones to weight weS like weG = 0.2 and weS = 0.8. The fuzzy positive association is tested with four different weights like wP = 0.0, wP = 0.1, wP = 0.5, and wP = 1.0.
Figure 3 presents the performance results for the models showing precision x recall curves. In these curves the best results occur when the precision values maintains high as the recall values increases. This indicates that most of relevant documents are retrieved and are presented in the top of the answer set. In the performed experiment only the best result is recorded for each model to keep the graphic understandable. The proposed model is represented by MO curves, the multi-relationship fuzzy concept network model by the CN curve and the Apache Lucene by LUC curves. In the curves legend, the KW means the use of just the entered keywords (without performing query expansion) and the numbers refer to the corresponding wP value when query expansion is considered. All the models showed a behavior tendency considering the wP variations.
As the proposed model and the Apache Lucene use the same query expansion method their performances have the same tendency. When considering just the user entered keywords the precision for lower recall values is high but it decreases fast as the recall values increase. When query expansion is performed the precision is high for low recall values and maintains around 45% for high recall values. The fuzzy concept network model presents high precision values for low recall values but as the recall values are higher it maintains the precision values around 25%. Comparing the three models results, the proposed model exhibits better performance (MO 0.1 curve) when knowledge is considered in the query expansion process. 5
This work presents an approach for improving document retrieval process considering a knowledge base composed of multiple related domain ontologies. Contrary to other approaches that consider the knowledge base composed of just one ontology, the proposed model explores knowledge expressed in multiple ontologies that, in some contexts, can be related to each other by causal, spatial or similarity relationships. To deal with the uncertainty and vagueness present in the knowledge the fuzzy set theory is used to express the relations between concepts of distinct ontologies. This knowledge is used in a novel method to expand the user query and to index the documents in the collection. Experimental results show that the proposed model achieves better performance when compared with other fuzzy information retrieval approach. When using the expansion method with the Apache Lucene search engine the results are also improved.
The knowledge organization and representation as ontologies is a growing
area. Many independent developed crisp ontologies, representing distinct
domains, are being proposed. The presented model offers an way where these
ontologies can be reused. Instead of developing one large ontology that encodes
multidisciplinary knowledge an alternative is to encode this knowledge as distinct
domain ontologies relating them in a next step. This allows distinct knowledge
groups to work in a independent way or to reuse already existent domain
ontologies. If the ontologies represent domains that can be related in some context
then just the positive fuzzy associations between the ontologies concepts have to
be constructed in order to reuse them in the model. An experiment considering
the ontologies themselves as fuzzy structures is presented in [