- (Biemann, 2005) ⇒ Chris Biemann. (2005). “Ontology Learning from Text: A Survey of Methods.” In: LDV Forum Journal, 20(2).
Subject Headings: Ontology Learning from Text.
After the vision of the Semantic Web was broadcasted at the turn of the millennium, ontology became a synonym for the solution to many problems concerning the fact that computers do not understand human language: if there were an ontology and every document were marked up with it and we had agents that would understand the markup, then computers would finally be able to process our queries in a really sophisticated way. Some years later, the success of Google shows us that the vision has not come true, being hampered by the incredible amount of extra work required for the intellectual encoding of semantic mark-up – as compared to simply uploading an HTML page. To alleviate this acquisition bottleneck, the field of ontology learning has since emerged as an important sub-field of ontology engineering.
It is widely accepted that ontologies can facilitate text understanding and automatic processing of textual resources. Moving from words to concepts not only mitigates data sparseness issues, but also promises appealing solutions to polysemy and homonymy by finding non-ambiguous concepts that may map to various realizations in – possibly ambiguous – words.
Numerous applications using lexical-semantic databases like WordNet (Miller, 1990) and its non-English counterparts, e.g. EuroWordNet (Vossen, 1997) or CoreNet (Choi and Bae, 2004) demonstrate the utility of semantic resources for natural language processing.
Learning semantic resources from text instead of manually creating them might be dangerous in terms of correctness, but has undeniable advantages: Creating resources for text processing from the texts to be processed will fit the semantic component neatly and directly to them, which will never be possible with general-purpose resources. Further, the cost per entry is greatly reduced, giving rise to much larger resources than an advocate of a manual approach could ever afford. On the other hand, none of the methods used today are good enough for creating semantic resources of any kind in a completely unsupervised fashion, albeit automatic methods can facilitate manual construction to a large extent.
The term ontology is understood in a variety of ways and has been used in philosophy for many centuries. In contrast, the notion of ontology in the field of computer science is younger – but almost used as inconsistently, when it comes to the details of the definition.
The intention of this essay is to give an overview of different methods that learn ontologies or ontology-like structures from unstructured text. Ontology learning from other sources, issues in description languages, ontology editors, ontology merging and ontology evolving transcend the scope of this article. Surveys on ontology learning from text and other sources can be found in Ding and Foo (2002) and Gómez-Pérez and Manzano-Macho (2003), for a survey of ontology learning from the Semantic Web perspective the reader is referred to Omelayenko (2001).
Another goal of this essay is to clarify the notion of the term ontology not by defining it once and for all, but to illustrate the correspondences and differences of its usage.
In the remainder of this section, the usage of ontology is illustrated very briefly in the field of philosophy as contrasted to computer science, where different types of ontologies can be identified.
In section 2, a variety of methods for learning ontologies from unstructured text sources are classified and explained on a conceptual level. Section 3 deals with the evaluation of automatically generated ontologies and section 4 concludes.
1.1 Ontology in philosophy
In philosophy, the term ontology refers to the study of existence. In this sense, the subject is already a central topic of Aristotle’s Categories and in all metaphysics. The term was introduced in the later Renaissance period, see Ritter and Gründer (1995), as “lat. philosophia de ente”. In the course of centuries, ontology was specified in different ways and covered various aspects of metaphysics. It was sometimes even used as a synonym for this field. Further, the distinction between ontology and theology was not at all times clear and began to emerge in the 17th century.
1.2 Ontologies in Computer Science
Ontology in [[[computer science]] is understood not as general as in philosophy, because the perception of ontologies is influenced by application-based thinking. But still ontologies in computer science aim at explaining the world(s), however, instead of embracing the whole picture, they only focus on what is called a domain. A domain is, so to speak, the world as perceived by an application. Example: The application of a fridge is to keep its interior cold and that is reached by a cooling mechanism which is triggered by a thermostat. So the domain of the fridge consists only of the mechanism and the thermostat, not of the food in the fridge, and can be expressed formally in a fridge ontology. Whenever the application of the fridge is extended, e.g. to illuminate the interior when the door is opened, the fridge ontology has to be changed to meet the new requirements. So much about the fridge world. In real applications, domains are much more complicated and cannot be overseen at a glance.
Ontologies in computer science are specifications of shared conceptualizations of a domain of interest that are shared by a group of people. Mostly, they build upon a hierarchical backbone and can be separated into two levels: upper ontologies and domain ontologies.
Upper ontologies (or foundation ontologies), which describe the most general entities, contain very generic specifications and serve as a foundation for specializations. Two well-known upper ontologies are SUMO (Pease and Niles, 2002) and CyC (Lenat, 1995). Typical entries in upper ontologies are e.g. “entity”, “object” and “situation”, which subsume a large number of more specific concepts. Learning these upper levels of ontologies from text seems a very tedious, if not impossible task: The connections as expressed by upper ontologies consist of general world knowledge that is rather not acquired by language and is not explicitly lexicalized in texts.
Domain ontologies, on the other hand, aim at describing a subject domain. Entities and relations of a specific domain are sometimes expressed directly in the texts belonging to it and can eventually be extracted. In this case, two facts are advantageous for learning the ontological structures from text: The more specialized the domain, the less is the influence of word sense ambiguity according to the “one sense per domain”-assumption in analogy to the “one sense per discourse”-assumption (Gale et al., 1993). Additionally, the less common-knowledge a fact is, the more likely it is to be mentioned in textual form.
In the following section, distinctions between different kinds of ontologies and other ways of categorizing the world are drawn.
|2005 OntologyLearningfromTextASurvey||Chris Biemann||Ontology Learning from Text: A Survey of Methods||2005|