MIsA: Multilingual “IsA” Extraction from Corpora

MIsA: Multilingual "IsA" Extraction from Corpora Stefano Faralli1, Els Lefever2, Simone Paolo Ponzetto1 1Data and Web Science, University of Mannheim, University of Mannheim {stefano, simone}@informatik.uni-mannheim.de 2LT3, Language and Translation Technology Team, Ghent University, Belgium els.lefever@ugent.be Abstract In this paper we present Multilingual IsA (MIsA), which is a collection of hypernymy relations in five languages (i.e., English, Spanish, French, Italian and Dutch) extracted from the corresponding full Wikipedia corpus. For each language, we first established a set of existing (viz. found in literature) or newly defined lexico-syntactic patterns. Similarly to WebIsADb, the resulting resource contains hypernymy relations represented as "tuples", as well as additional information such as provenance (Gil and Groth, 2011), context of the extraction, etc. To measure the precision of the patterns, we performed a manual assessment of the quality of the extracted relations and an error analysis. In addition, we release the software developed for the extraction of the hypernym tuples. Keywords: multilinguality, Hearst patterns, hypernym extraction, framework. 1. Introduction Hypernymy relations represent the relationship between a generic term (hypernym) and a specific instance of it (hy- ponym). These relations play a key role for many Natural Language Processing (NLP) tasks, e.g. ontology learning, automatically building or extending knowledge bases, or word sense disambiguation and induction. In fact, hyper- nymy relations may provide the basis for the construction of more complex structures such as taxonomies, or be used as effective background knowledge for many word under- standing applications. In the past, many different methods have been developed for hypernym extraction, ranging from simple lexical pat- terns (Hearst, 1992; Oakes, 2005) to statistical and machine learning techniques (Dolan et al., 1993; Caraballo, 1999; Agirre et al., 2000; Ritter et al., 2009), to name a few. Snow et al. (2004) first search sentences that contain two terms that are known to be in a taxonomic relation (term pairs are taken from WordNet (Miller et al., 1990)), then parse the sentences, and automatically learn patterns from the parse trees. Finally, they train a hypernym classifier based on these features. Lexico-syntactic patterns are gen- erated for each sentence relating a term to its hypernym, and a dependency parser is used to represent them. For the ontology learning task, Velardi et al. (2013) induce taxonomies from scratch by extracting hypernyms from a domain corpus and the Web. Definitional sentences such as “lion is a dangerous animal” (where “animal” is the hy- pernym of “lion”) are recognized by the Word Class Lat- tices classifier (Navigli and Velardi, 2010) trained on a large set of Wikipedia definitions. Kozareva and Hovy (2010) induce a taxonomy using a par- ticular kind of Hearst-like (Hearst, 1992) lexico-syntactic patterns, i.e. so-called Doubly Anchored Patterns (DAP). The hypernymy relation extraction consists of two phases. First, the authors bootstrap the terminology harvesting with DAP of the kind “animals such as lions and *”, so it is possible to discover new terms such as “cats”. Next, for each pair of terms in the discovered terminology, e.g. (“li- ons”,“cats”), they automatically create a …

Content truncated to 500 words