Learning domain taxonomies: the TaxoLine approach

• DOI: https://doi.org/10.1108/IJWIS-04-2017-0024
• Pages: 281-301
• Date: 21 August 2017
• Published date: 21 August 2017
• Author: Omar El Idrissi Esserhrouchni,Bouchra Frikh,Brahim Ouhbi,Ismail Khalil Ibrahim
• Subject Matter: Information & knowledge management,Information & communications technology,Information systems,Library & information science,Information behaviour & retrieval,Metadata,Internet

Learning domain taxonomies:

the TaxoLine approach

Omar El Idrissi Esserhrouchni

Université Moulay Ismail Ecole Nationale Supérieure d’Arts et Metiers

(ENSAM), Morocco

Bouchra Frikh

Université Sidi Mohamed Ben Abdallah Ecole Supérieur de Technologie

(ESTF)–LTTI Lab Fès, Morocco

Brahim Ouhbi

Université Moulay Ismail Ecole Nationale Supérieure d’Arts et Metiers

(ENSAM), Meknès, Morocco, and

Ismail Khalil Ibrahim

Department of Telecooperation, Johannes Kepler University Linz,

Linz, Austria

Abstract

Purpose –The aim of this paper is to present an online framework for building a domain taxonomy, called

TaxoLine, from Web documents automatically.

Design/methodology/approach –TaxoLine proposes an innovative methodology that combines

frequency and conditional mutual information to improve the quality of the domain taxonomy. The system

also includes a set of mechanisms that improve the execution time needed to build the ontology.

Findings –The performance of the TaxoLine framework was applied to nine different nancial corpora.

The generated taxonomies are evaluated against a gold-standard ontology and are compared to

state-of-the-art ontology learning methods.

Originality/value –The experimental results show that TaxoLine produces high precision and recall for

both concept and relation extraction than well-known ontology learning algorithms. Furthermore, it also

shows promising results in terms of execution time needed to build the domain taxonomy.

Keywords Chir-statistic, Conditional mutual information, Domain taxonomy, Financial ontology,

Financial taxonomy, Ontology learning

Paper type Research paper

1. Introduction and motivation

The proliferation of the internet and Web content-producing services have allowed the

creation of a huge amount of unstructured data on the Web. These data represent an

impressive source of information for knowledge extraction. However, to infer knowledge and

extract semantics from this information, it is necessary to represent them in a structured way

to make them machine-processable. This could be performed by generating domain

taxonomies from Web documents corpora (Maedche et al., 2003).

A taxonomy is dened as a specic form of an ontology, which is a formal and an explicit

specication of a shared conceptualization (Guber, 1993). Ontologies are more complex than

taxonomies. They consist of multiple types of semantic relations between concepts,

including part-of and other domain-specic relations. A taxonomy is a concept hierarchy

that includes only the hierarchical parent-child relations. Taxonomies can be represented in

The current issue and full text archive of this journal is available on Emerald Insight at:

www.emeraldinsight.com/1744-0084.htm

Learning

domain

taxonomies

281

Received 24 April 2017

Revised 24 April 2017

Accepted 1 May 2017

InternationalJournal of Web

InformationSystems

Vol.13 No. 3, 2017

pp.281-301

©Emerald Publishing Limited

1744-0084

DOI 10.1108/IJWIS-04-2017-0024

a declarative form by using semantic markup languages like RDF and OWL. They have been

used in various elds such as e-commerce (Domingue et al., 2003), medicine (Wen et al., 2014)

and biomedical informatics (Paul and Maji, 2014).

However, manual construction of a domain taxonomy remains a costly task and

time-consuming process (Dhingra and Bhatia, 2012). It requires an extended knowledge of

the studied domain (expert). Therefore, to reduce the effort of building domain taxonomies,

several approaches that target building domain taxonomies, automatically or

semi-automatically, have been developed using different techniques (Cimiano et al., 2004;

Hearst, 1992;Meijer et al., 2014;Novalija et al., 2011).

Nevertheless, according to Buitelaar et al. (2005), the process of learning a domain

taxonomy is based on many sub-tasks and issues that have to be addressed. One of the main

issues is the data sparseness: usually, it is difcult to obtain enough data that cover the

domain of interest thoroughly. However, learning concept hierarchy requires a large amount

of data to build an appropriate domain taxonomy that covers the most important concepts.

To overcome this problem, we use the Web as a source of knowledge.

The next issue is whether the use of frequency analysis only should select the most

relevant concept and should identify precisely their taxonomic relationships. We

hypothesize that a domain taxonomy which is constructed with an algorithm that combines

statistic frequency and semantic similarity measure will extract the taxonomy more

precisely than an algorithm based on frequency only.

Another issue that, to the best of our knowledge, was never discussed in earlier works and

that needs to be addressed, is the execution time constraint. The execution time of an

algorithm is a very important parameter to evaluate its applicability. We believe that an

effective algorithm is the one which extracts concept hierarchy and builds the taxonomy

accurately and promptly.

Frikh et al. (2011) presented HCHIRSIM, an algorithm to build domain taxonomy

automatically from Web documents. The approach combines both a frequency statistic and

a similarity measure based on mutual information to build a domain taxonomy. The

algorithm was evaluated in the cancer domain against the well-known MESH ontology and

showed promising results. However, the approach still has some limitations in nding

hyponyms between concepts. Indeed, the mutual information used to measure the similarity

between two concepts does not take into consideration the context of their parent. For

instance, the term “Bank” means a nancial institution in the nancial context and a ight

maneuver in the aviation context. Nevertheless, both taxonomic relations will be extracted

by the algorithm and will be integrated in the result taxonomy, even if the studied domain is

nance or aviation. Thus, the accuracy of the result taxonomy is decreased.

In this work, we present a novel system, called TaxoLine[1] which overcomes this

limitation by introducing a new measure based on conditional mutual information.

Therefore, to identify the dependency between two concepts, the proposed algorithm

measures the similarity between them on the basis of the presence of their parent concept.

On the basis of the tests carried out, TaxoLine is more efcient and fully optimized

because it introduces a new measure based on the conditional mutual information (CMI), and

compared to benchmark algorithms, it perfectly improves the execution time of the

taxonomy learning process by using indexation method, optimized data structures and

threads in the implementation of the algorithm. To the best of our knowledge, earlier

research has rarely considered the execution time (time needed for the algorithm to construct

the taxonomy) as a property in the evaluation process. The purpose of our new algorithm is

to make the system fast enough to be hosted on the Web and used online.

IJWIS

13,3

282