A Modular Framework to Learn Seed Ontologies from Text

Introduction

In the last years, there has been a considerable increase in research on knowledge-based systems, especially in the context of the Semantic Web. However these systems, as they aim at something more than just supplying trivial functionalities, suffer in their development process of the so-called knowledge acquisition bottleneck: creating large, usable, expandable and valid representations of semantics about a specific domain of interest, i.e., ontologies, represents the most time-consuming task of the whole project.

One of the main reasons for this knowledge acquisition bottleneck is that these formal representations, representing semantics previously unknown to machines, need to be manually annotated by domain experts. In this ontology building process, two different kinds of expertise are usually required: the knowledge of the domain that has to be described, and the ability to encode the ontology in a machine interpretable, i.e., computational, format. Unfortunately, satisfying both of these requirements is, in most cases, not trivial (if not impossible). On the one hand, domain knowledge might be very specific and known only to a small community of practice usually not in the realm of ontology writers; on the other hand, the encoding process does not only depend on standards and tools, but also on the specific context in which this knowledge has to be used.

Easing this task means either making semantic technologies more accessible to domain experts or providing ontology experts with structured information about a domain. The work presented in this chapter is mainly focused on the second approach and it aims at providing an alternative to the manual generation of ontologies through the automatic extraction of candidate concepts and relationships from a set of documents. The generated seed ontology is just an initial and approximate representation of the domain knowledge, and it obviously needs to be further modified and expanded, but it considerably reduces the time required for the overall formalization of the domain knowledge from scratch.

There is already plenty of information available on the Internet (and in other more reliable digital libraries) which could be used to teach a machine about virtually any domain of knowledge. This information is stored as collections of Web pages, large document corpora, databases, and so on. These repositories, however, cannot be directly consumed by a machine as they contain no structured information according to any standard model for knowledge representation. In this scenario, the main challenge we are interested in is the so-called Ontology Learning from Text. Free text is a valuable source of unstructured knowledge, and this requires researchers to adopt original heuristics in order to extract structured semantics from it. These heuristics often return inaccurate results, and have to be modified and validated by experts of the domain, but they can be exploited to bootstrap the whole ontology learning process.

Our work aims at providing a modular, semi-automatic framework that allows its users to apply different heuristics, (possibly) combine them, and finally measure their accuracy. The framework splits the process of ontology learning from text in well-defined steps and relies on different techniques at each stage. This provides some additional benefits to the whole process; for instance, it allows the use of the system with different combinations of algorithms, or to access any useful information generated at intermediate stages. The outcome of this work is a tool called Extraction, which allows the identification of main concepts and their basic relationships from a corpus of free text documents about a specific domain of knowledge.

This chapter has two purposes. On the one hand, we want to describe our modular architecture for the semi-automatic extraction of ontologies from text. This part should be especially useful for anyone trying to develop a similar system, as we provide insights into the building process and show the main issues we had to face, together with the choices we made to solve them. On the other hand, we review existing approaches for the extraction of relationships between terms, we evaluate them in the context of our framework, and we introduce a novel approach aimed at identifying new types of relationships, Action and Affection, which are distinct from classical subsumption.