Instance-Based Ontology Population Exploiting Named-Entity Substitution

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 265–272 Manchester, August 2008 Instance-Based Ontology Population Exploiting Named-Entity Substitution Claudio Giuliano Fondazione Bruno Kessler Trento, Italy giuliano@fbk.eu Alfio Gliozzo Laboratory for Applied Ontology Italian National Research Council Rome, Italy alfio.gliozzo@cnr.istc.it Abstract We present an approach to ontology popu- lation based on a lexical substitution tech- nique. It consists in estimating the plausi- bility of sentences where the named entity to be classified is substituted with the ones contained in the training data, in our case, a partially populated ontology. Plausibility is estimated by using Web data, while the classification algorithm is instance-based. We evaluated our method on two different ontology population tasks. Experiments show that our solution is effective, out- performing existing methods, and it can be applied to practical ontology population problems. 1 Introduction Semantic Web and knowledge management appli- cations require to populate the concepts of their domain ontologies with individuals and find their relationships from various data sources, including databases and natural language texts. As the ex- tensional part of an ontology (the ABox) is often manually populated, this activity can be very time- consuming, requiring considerable human effort. The development of automatic techniques for on- tology population is then a crucial research area. Natural language processing techniques are natu- ral candidates to solve this problem as most of the data contained in the Web and in the companies’ intranets is free text. Information extraction (IE) is commonly employed to (semi-) automate such a task. c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. Current state-of-the-art IE systems are mostly based on general purpose supervised machine learning techniques (e.g., kernel methods). How- ever, supervised systems achieve acceptable accu- racy only if they are supplied with a sufficiently large amount of training data, usually consisting of manually annotated texts. Consequently, they can be only used to populate top-level concepts of on- tologies (e.g., people, locations, organizations). In fact, when the number of subclasses increases the number of annotated documents required to find sufficient positive examples for all subclasses be- comes too large to be practical. As domain ontolo- gies usually contain hundreds of concepts arranged in deep class/subclass hierarchies, alternative tech- niques have to be found to recognize fine-grained distinctions (e.g., to categorize people as scientists and scientists as physicists, mathematicians, biol- ogists, etc.). In this paper, we present an approach to the clas- sification of named entities into fine-grained onto- logical categories based on a method successfully employed in lexical substitution.1 In particular, we predict the fine-grained category of a named en- tity, previously recognized, by simply estimating the plausibility of sentences where the entity to be classified is substituted with the ones contained in the training data, in our case, a partially populated ontology. In most of the cases, ontologies are partially populated during the development phase and af- ter that the annotation cost is practically negligi- ble, making this method highly attractive in many …

Content truncated to 500 words