Unsupervised Knowledge Extraction for Taxonomies of Concepts from Wikipedia

International Conference RANLP 2009 - Borovets, Bulgaria, pages 28–32 Unsupervised Knowledge Extraction for Taxonomies of Concepts from Wikipedia Eduard Barbu Center for Mind/Brain Sciences Rovereto Trento, Italy eduard.barbu@unitn.it Massimo Poesio Center for Mind/Brain Sciences Rovereto Trento, Italy massimo.poesio@unitn.it Abstract A novel method for unsupervised acquisition of knowledge for taxonomies of concepts from raw Wikipedia text is presented. We assume that the concepts classified under the same node in a taxonomy are described in a comparable way in Wikipedia. The concepts in 6 tax- onomies extracted from WordNet are mapped onto Wikipedia pages and the lexico-syntactic patterns describing semantic structures express- ing relevant knowledge for the concepts are au- tomatically learnt. Keywords wikipedia, unsupervised knowledge acquisition, taxonomy 1 Introduction A crucial phase in ontology acquisition from text is the extraction of relevant knowledge for ontology con- cepts, the focus of the current work. Our framework extracts in an unsupervised way knowledge for a set of concepts hierarchically ordered. For example, for the concept bewick’s swan, one of the concepts in bird taxonomy, some extracted properties are: have few natural predator 1, live in water, is a small Hol- arctic swan. From a logical/ontological point of view the extracted knowledge can be classified as: quanti- fier restrictions (e. g. most birds build nests), parts of the instances of the concepts in the taxonomy (e.g. small head and long thick mane for the concept shet- land pony), alternative classification of the concepts in the taxonomy (herd animal and social creature for the concept horse), etc. The knowledge relevant for concepts can be auto- matically extracted from a variety of sources: dictio- naries, databases, corpora, web directories and others. Recently, Wikipedia drew the attention of various re- search groups as a goldmine resource for information retrieval [3], information extraction [9] and ontology building [8]. There are some characteristics that make Wikipedia an appropriate resource for information extraction. Firstly, its coverage is impressive: the English Wikipedia has almost three million articles currently 1 In this paper the concepts will be typed in bold and the properties in italics maintained and updated by thousands of voluntary contributors, thus surpassing any other encyclopedia in history. Secondly, the style of writing Wikipedia articles is more homogeneous than the mixed bag of styles one encounters in general corpora or in unre- stricted text found on the web. Thirdly, Wikipedia has a large network of links, categories and info-boxes allowing a combination of techniques for information extraction. This paper introduces a novel method for acquisition of knowledge for taxonomies of concepts from the raw Wikipedia text. We assume that similar concepts (i.e. those classified under the same node in a taxonomy) are described in a comparable way in Wikipedia. More precisely, we suppose that the relevant knowledge of these similar concepts is expressed using equivalent surface patterns. The learning process starts with the generation of concept hierarchies from WordNet. The concepts in each hierarchy are mapped onto Wikipedia pages and the knowledge appropriate to the concepts is automatically extracted at …

Content truncated to 500 words