Supporting the discovery and labeling of non-taxonomic relationships in ontology learning

PARNT: A Statistic based Approach to Extract Non-Taxonomic Relationships of Ontologies from Text

“…For example, Maedche and Staab [2] consider the existence of two instances of ontology concepts in a sentence as evidence that they are non-taxonomically related. For Villaverde et al [12], a relationship is identified by the presence of two ontology concepts in the same sentence with a verb between them.…”

“…Athough it performs the proposed functionality of suggesting the adequate hierarchical level for the non-taxonomic relationships, it is hardly applicable in practice. Villaverde et al [12] use, in the refinement phase, an algorithm for the extraction of association rules [22] that recommends nontaxonomic relationships including its corresponding verbal labels. However, in practice this solution leads to a decrease in effectiveness when compared to algorithms that suggest a list of labels for each pair of concepts like bag of labels (section IV) or even to those that do not suggest any label, like the frequency of tuples composed of two concepts (<c 1 , c 2 >).…”

“…Villaverde et al [12] proposed a LNTRO technique that uses part of speech tagging in the corpus annotation phase in order to identify the verbs. In the extraction of relationships a tuple in the form <c 1 , v, c 2 > (c 1 and c 2 are ontology concepts and v is a verb) is generated for each occurrence of two consecutive concepts with a verb between them.…”

“…Some techniques have already been proposed for Learning Non-Taxonomic Relationships of Ontologies (LNTRO) [1] [2] [5] [12] [21] [25]. All of them use Natural Language Processing (NLP) techniques [10] [18] to annotate the corpus with the information needed for subsequent processing, as well as Machine Learning (ML) [4] [24] to refine the relationships that result from the previous phases.…”

“…All of them use Natural Language Processing (NLP) techniques [10] [18] to annotate the corpus with the information needed for subsequent processing, as well as Machine Learning (ML) [4] [24] to refine the relationships that result from the previous phases. However, most of the techniques that use ML [2] [12], have limited application since they require corpora with specific characteristics, namely having a great amount of verb phrases composed of only one word or texts structured in titles and text body, as it will be explained in section III.…”