On Automated Hyperonym Hierarchy Construction Using an Internet Search Engine

On Automated Hypernym Hierarchy Construction Using an Internet Search Engine⋆ Sergey Afonin Institute of Mechanics, Moscow State University, 119192, Michurinskij av., 1, Moscow, Russian Federation serg@msu.ru Abstract. In this paper we propose an approach for automatic construction of concept hier- archies from the snippets returned by Internet search engines using a number of well known techniques. We use surface lexical patterns to construct a set of candidate hypernyms of a given term and additional filtering that is based on both lexical patterns and distributional analysis. Preliminary experimental results for real life English examples are presented. Keywords: ontology learning, concept hierarchy, hypernyms, lexical patterns. 1 Introduction Ontology is an ambiguous notion. In philosophy it goes back to Ancient Greeks and refers to the study of nature, or more generally, to the study of what might exist (Smith, 2004). In the context of computer science, terminological, information and knowledge anthologies (van Heijst et al., 1997) were introduced. Later two ontologies deal with concepts, while terminological ontologies are focused on natural language terms that may or may not be directly mapped into concepts. In contrast to information ontologies, knowledge ontologies use some formal languages to represent sematic relations between concepts. Terminological ontologies, such as WordNet, contain information about hyponym-hypernym (subtype-supertype, or IS-A relation), synonymy, and other semantic relations between terms. Such information can be used in variety of natural language processing applications, including query reformulation (Jones et al., 2006; Chirita et al., 2007), text summarization (Dang et al., 2008), categorization (Li et al., 2009), query answering (Lopez et al., 2007), and many others. Unfortunately, good quality ontologies are mostly manually created and they exist for a limited number of domains. Moreover, it is difficult to support an ontology for rapidly changing domains when new concepts or semantic relations frequently appear, so ontology construction and updat- ing become a bottleneck for many applications. A possible solution to the above problem is an automatic or semi-automatic ontology learning. There is a lot of work has been done in the area of ontology learning during past two decades. While there exist many types of semantic relations between concepts and many different sources of information (e.g. databases, structured text, dictionaries) the problem of concept hierarchy learning from unstructured texts may be considered as the main topic in the area. An importance of this problem can justified by the fact that concept hierarchy forms a core of every ontology and that unstructured text is the most frequent data format. Automated ontology learning from text is a challenging task. It is well known that the set of concepts and terminology varies between different languages and cultures. For example, in many northen languages there exist a lot of terms describing specific conditions of snow, that ⋆ This work was supported in part by the Russian Foundation of Basic Research under grant 09-07-00366a. Copyright 2010 by Sergey Afonin PACLIC 24 Proceedings 341 can not be easily translated into other languages. Even in the same language some terms can be treated differently. …

Content truncated to 500 words