Ontology

Definition of Ontology reference at ` - A Study Investigating Typical Concepts and Guidelines for Ontology Building, 2015 - https://www.igi-global.com/dictionary/ontology/21117

The domain knowledge in ontologies can be formalized using 5 components A translation approach to portable ontology specification

Classes: Set of classes (or concepts) that belong to the ontology. They may contain individuals (or instances), other classes, or a combination of both with their corresponding attributes.
Relations: These define interrelations between two or several classes (object properties) or a concept to a data type (data type properties).
Functions: This is a special case of relations.
Axioms: These are used to impose constraints on the values of classes or instances. Axioms represent expressions in ontology (logical statement) and are always true when used inside the ontology.
Instances: These represent the objects, elements or individuals of an ontology.

Concept

Based on [1], There is no unified definition of concept. A concept can be recognized as the relevant terms that shared the same meaning. Psychologically, it defnied as an idea shared in common in the minds of people who use these terms. It can also seen as a logical contruct, for example, a synset containing a set of words which can be exchanged for each other.

From [2]. A concept should include: (1) an intention of this concept; (2) a set a concept instances (ie. its extension); (3) a set of linguistic realizations (multiligual terms that express this concept). However, most of academic research regard concepts as clusters of related terms.

Concept Formation

Existing Approaches 1. Clustering over word embedding 2. Using Graph-based method and treating this problem as community discovery 3. Using Topic model-based method and viewing this problem as topic mining

Taxonomy Extraction

Existing Approaches 1. Hearst Patterns: By using a set of predefined lexico-syntactic patterns. This method can get good precision, but low recall.

1 2	Vehicles such as cars, trucks and bikes Swimming, running or other activities

2. Clustering:

Agglomerative clustering and Divisive clustering:
Terms are represented by a dense vector. Then a clustering algorithm is ran over all term vectors, therefore build up a hierarchy of clusters, where each cluster is regarded as a concept.

Conceptual Clustering: A lattice of terms is built by investigate the overlap of descriptive attributes (a document describe the term) between two represented terms.

Cases of conceptual clustering:
- An application of inductive concept analysis to construction of domain-specific ontologies, 2003
- Learning concept hierarchies from text corpora using formal concept analysis, 2005

Document-based subsumption: Term \(t_1\) subsumes \(t_2\) (ie. the relation of is-a(\(t_2\), \(t_1\))) if \(t_1\) appears in all the documents where \(t_2\) shows up. \(P(x|y)=\frac{freq_d(x,y)}{freq_d(y)}\)
Phrase Analysis: The internal structure of noun phrases can be used to deduce taxonomic relations. For example, additional modifiers added to the front of a noun usually is a subclass of the class denoted by the noun (focal epilepsy is subclass of epilepsy). Or terms occuring to the left of a term are subclasses of this term.

References are: - A Prot ́eg ́e plug-in for ontology extrac- tion from text based on linguistic analysis, 2004 - Web-scale taxonomy learning., 2005

Ontology Learning

Machine Learning Algorithm for Ontology Learning (OL)

Algorithm	Generic Use	Usage in OL
Association rule discovery	Discover interesting itermsets from transactions	Discover interesting association between terms
(Hierarchical) Clustering	Discover group in data	Clustering terms
SVM, NB, KNN	Prediction	Classification of new concepts into existing hierarchy
Inductive Logic Programming	Induction of rules from data (supervised)	Discovery of new concepts from extensional data
Conceptual Clustering (Formal Concept Analysis)	Concept discovery (extension and intention)	Learning concepts and concept hierarchy