The Semi-automatic Expansion of Existing Terminological Ontologies using Knowledge Patterns Discovered on the WWW: An Implementation and Evaluation

Jakob Halskov

    Publikation: Bog/antologi/afhandling/rapportPh.d.-afhandling

    3 Downloads (Pure)

    Abstrakt

    The research object of this thesis is the so-called knowledge patterns and their usefulness in automatically extracting specic semantic relations from unannotated and uncategorized text on the WWW so as to facilitate semi-automatic updating and extension of existing ontological and terminological resources. The main contribution of the thesis is the implementation of a com- plete ontology extension framework called WWW2REL which is 100% based on a knowledge-poor, domain-independent processing of WWW text snippets and includes the three stages of pattern discovery, pattern ltering and relation instance ranking. Unlike most comparable systems WWW2REL is special in that it is both highly portable, can be applied to any semantic relation type and operates directly on uncategorized WWW text snippets. The system is tested on the biomedical UMLS Metathesaurus for four dierent relation types and manually evaluated by four domain experts. It is demonstrated that high precision in the task of knowledge discovery from a noisy text source can be achieved using a very simple instance relevance measure and two ranking heuristics. In contrast, many comparable systems operate on richly annotated academic text and tend to apply heuristics which are custom-tailored to a specic domain and/or relation type. When selecting the overall best ranking scheme, average system performance across all four relation types ranges between 70% to 65% of the maximum possible F-score by top 10 and top 50 relation instances, respectively. Finally, the thesis experiments also examine the portability of individ- ual knowledge patterns and of the ranking heuristics. It is concluded that synonymy KPs are the most domain independent closely followed by ISA KPs, whereas patterns for "may_prevent" and especially "induces" are more dependent on the domain. Empirical experiments also suggest that a ranking heuristic which penalizes relation instances whose arguments occur frequently in a general language corpus can be highly eective, but may need to be adapted to the domain in question.
    OriginalsprogEngelsk
    UdgivelsesstedFrederiksberg
    ForlagCopenhagen Business School [Phd]
    Antal sider289
    ISBN (Trykt)9788759383377
    StatusUdgivet - 2007
    NavnPhD series
    Nummer28.2007
    ISSN0906-6934

    Emneord

    • Ph.d.-afhandlinger
    • Datalingvistik
    • Terminologi
    • Terminologiekstraktion
    • Internet
    • Automatisk fremfinding
    • Ontologier
    • WWW2REL

    Citationsformater