INRIASAC: Simple Hypernym Extraction Methods

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 911–914, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics INRIASAC: Simple Hypernym Extraction Methods Gregory Grefenstette Inria 1 rue Honoré d'Estienne d'Orves 91120 Palaiseau, France Gregory.grefenstette@inria.fr Abstract For information retrieval, it is useful to classi- fy documents using a hierarchy of terms from a domain. One problem is that, for many do- mains, hierarchies of terms are not available. The task 17 of SemEval 2015 addresses the problem of structuring a set of terms from a given domain into a taxonomy without manu- al intervention. Here we present some simple taxonomy structuring techniques, such as term overlap and document and sentence co- occurrence in large quantities of text (English Wikipedia) to produce hypernym pairs for the eight domain lists supplied by the task organ- izers. Our submission ranked first in this 2015 benchmark, which suggests that overly com- plicated methods might need to be adapted to individual domains. We describe our generic techniques and present an initial evaluation of results. 1 Introduction This paper describes two simple hypernym extrac- tion methods, given a list of domain terms and a large amount of text divided into documents. Task 17 of the 2015 Semeval campaign (Bordea et al., 2015) consists in structuring a flat list of pre- identified domain terms into a list of hypernym pairs. Task organizers provide two lists of terms for each of four domains: equipment, food, chemi- cal, science, one extracted from WordNet and one from an unknown source. Participants in the task were allowed to use any resource (except existing taxonomies) to automatically transform the lists of terms into lists of pairs of terms, the first term be- ing a hyponym of the more general second term. For example, if the words airship and blimp were included in the lists of terms for a domain, the system was expected to return lines such as: blimp airship The task organizers provided training data from the domains of Artificial Intelligence, vehicles and plants, different from the test domains. The train- ing data consisted in term lists (for plants), and term lists and lists of hypernyms (for AI and for vehicles). We examined these files to get an under- standing of the task but did not exploit them. We used the English text of Wikipedia (download- ed from http://dumps.wikimedia.org on August 13, 2014) as our only resource for discovering these relations. We extracted only the text of each arti- cle, ignoring titles, section headings, categories, infoboxes, or other meta-information present in the article. We recognized task terms in these articles and gathered statistics on document and sentence co-occurrence between domain terms, as well as term frequency. To recognize hypernyms, we used term inclusion (explained in section 3.1 below) and co-occurrence statistics (see section 3.2) to decide whether two terms were possibly in a hypernym relation, and document frequency to chose which term was the hypernym. Our submission ranked first in the SemEval 2015 task 17 benchmark. …

Content truncated to 500 words