6 Automated Geographical Information Fusion
113
in data) is it possible to begin to determine how concepts are actually used. Second,
extensional information forms a rich source of examples that can be used as the basis
for automated pattern recognition techniques.
Recognizing the importance of instance-level information, an increasing number
of researchers have turned to extensional approaches, including the following.
• The SemInt system clusters patterns in instance-level information, and uses these
clusters to train a neural network to identify intensional relationships [43, 44].
• Doan and collaborators [17, 18] and the Autoplex system [5] use Bayesian machine learning techniques on instance-level information to identify intensional
relationships.
• The Clio [49] and iMAP [16] systems search for filters that relate sets of
instance-level information within a database. These filters are then used to infer intensional relationships.
• He and Chang make use of patterns of co-occurrence of related attributes for
Web pages [36]. The positive correlation between related attributes, along with
an expected negative correlation between synonyms, is used to automatically
infer semantic mappings between attributes within a domain.
Fundamentally, all these extensional approaches apply different forms of inductive
inference: they use the structure and patterns in instance-level information to infer
semantic relationships. An inherent limitation of using inductive inference is that it
is unreliable. In many of the extensional approaches outlined above unreliability is
combated using probabilistic techniques (such as Bayesian probability). We return
to the topic of reasoning reliability in Sect. 6.4.
6.2.3 Geographical Information Fusion
Research into geographical information fusion mirrors the more general approaches
to information fusion cited above. Fonseca and coauthors have published a series of
papers on the so-called ontology-driven GIS [25–27]. This work aims to augment
conventional GIS with formal representations of geographical ontologies, leading to
tools that enable improved ontology-based information integration. A wide variety of
related work has addressed the issue of geographical information integration from a
similar perspective (e.g. [2, 7, 15, 58, 63]). In common with the research presented in
Sect. 6.2.1, such research focuses on the integration itself, but assumes the semantic
relationships between different ontologies are already known.
A relatively small amount of work has begun to provide tools for geographical
ontology alignment. Most of this work adopts an intensional approach. Kavouras,
Kokla, and coauthors use FCA as the basis for their approach to geographical ontology alignment [37, 38, 41]. Manoah et al. applied the intensional machine learning
techniques discussed in Sect. 6.2.2 to geographical data [46]. Duckham and Worboys
have investigated using description logics [20] and a formal algebraic approach [71]
to ontology alignment. Because of the diversity of geographical terms and concepts,
this work is at best semi-automated, and still requires human domain experts at critical stages in the alignment process.
113
in data) is it possible to begin to determine how concepts are actually used. Second,
extensional information forms a rich source of examples that can be used as the basis
for automated pattern recognition techniques.
Recognizing the importance of instance-level information, an increasing number
of researchers have turned to extensional approaches, including the following.
• The SemInt system clusters patterns in instance-level information, and uses these
clusters to train a neural network to identify intensional relationships [43, 44].
• Doan and collaborators [17, 18] and the Autoplex system [5] use Bayesian machine learning techniques on instance-level information to identify intensional
relationships.
• The Clio [49] and iMAP [16] systems search for filters that relate sets of
instance-level information within a database. These filters are then used to infer intensional relationships.
• He and Chang make use of patterns of co-occurrence of related attributes for
Web pages [36]. The positive correlation between related attributes, along with
an expected negative correlation between synonyms, is used to automatically
infer semantic mappings between attributes within a domain.
Fundamentally, all these extensional approaches apply different forms of inductive
inference: they use the structure and patterns in instance-level information to infer
semantic relationships. An inherent limitation of using inductive inference is that it
is unreliable. In many of the extensional approaches outlined above unreliability is
combated using probabilistic techniques (such as Bayesian probability). We return
to the topic of reasoning reliability in Sect. 6.4.
6.2.3 Geographical Information Fusion
Research into geographical information fusion mirrors the more general approaches
to information fusion cited above. Fonseca and coauthors have published a series of
papers on the so-called ontology-driven GIS [25–27]. This work aims to augment
conventional GIS with formal representations of geographical ontologies, leading to
tools that enable improved ontology-based information integration. A wide variety of
related work has addressed the issue of geographical information integration from a
similar perspective (e.g. [2, 7, 15, 58, 63]). In common with the research presented in
Sect. 6.2.1, such research focuses on the integration itself, but assumes the semantic
relationships between different ontologies are already known.
A relatively small amount of work has begun to provide tools for geographical
ontology alignment. Most of this work adopts an intensional approach. Kavouras,
Kokla, and coauthors use FCA as the basis for their approach to geographical ontology alignment [37, 38, 41]. Manoah et al. applied the intensional machine learning
techniques discussed in Sect. 6.2.2 to geographical data [46]. Duckham and Worboys
have investigated using description logics [20] and a formal algebraic approach [71]
to ontology alignment. Because of the diversity of geographical terms and concepts,
this work is at best semi-automated, and still requires human domain experts at critical stages in the alignment process.