A Combined Pattern-based and Distributional Approach for Automatic Hypernym Detection in Dutch.

Proceedings of Recent Advances in Natural Language Processing, pages 593–600, Hissar, Bulgaria, 7-13 September 2013. A Combined Pattern-based and Distributional Approach for Automatic Hypernym Detection in Dutch Gwendolijn Schropp LT3, Language and Translation Technology Team University College Ghent Gent, Belgium gwendolijn@gmail.com Els Lefever, V´eronique Hoste LT3, Language and Translation Technology Team Ghent University Gent, Belgium Els.Lefever@ugent.be Veronique.Hoste@ugent.be Abstract This paper proposes a two-step approach to find hypernym relations between pairs of noun phrases in Dutch text. We first apply a pattern-based approach that com- bines lexical and shallow syntactic infor- mation to extract a list of candidate hy- pernym pairs from the input text. In a second step, distributional similarity infor- mation is used to filter the obtained list of candidate pairs. Evaluation of the sys- tem shows encouraging results and reveals that the distributional information partic- ularly helps to improve the precision for context dependent hypernym pairs. The proposed hypernym module is considered an important step in building a semantic structure for automatically extracted ter- minology. As our approach does not re- quire external lexical resources, it can be applied for any given Dutch input text and is particularly well suited for domain and user specific text. 1 Introduction Recent work in knowledge-rich NLP tasks such as information retrieval, question answering, textual entailment and sentiment analysis have revealed a need for more structured data, where concepts are stored together with the semantic relation- ships that exist between these concepts and their corresponding surface forms. Structured lexical- semantic databases such as WordNet (Miller et al., 1990) or EuroWordNet (Vossen, 1998) have been deployed for a wide range of NLP tasks, but suffer from a number of shortcomings. Firstly, these manually crafted resources are very labour-intensive and costly to create. Secondly, existing lexical inventories contain more general vocabulary and have by consequence a low cover- age for domain-specific terms. As a consequence, researchers have started to investigate how seman- tic resources such as ontologies can be learned from text instead of being created manually. For an overview we refer to (Biemann, 2005). In this paper, we focus on the detection of hyper- nym relations between nouns and noun phrases. Automatic extraction of nouns or noun phrases which are semantically related has been suc- cessfully achieved in prior research, for example using coordination and co-occurrence information (Oh et al., 2009; Cederberg and Widdows, 2003; Roark and Charniak, 1998; Widdows and Borow, 2002). However, automatically distinguishing ex- actly which semantic relationship exists between them is not that straightforward. One of these semantic relationships is the hypernym relation which can be seen as a set-subset relation. In the literature the following description is adopted the most: a(n) NP0 is a (kind of) NP1; where NP1 is the hypernym of NP0 (which is in turn the hyponym) and the relationship is reflexive and transitive but not symmetric (Miller et al., 1990; Hearst, 1992). Note the subtle difference with meronymy (Girju et al., 2003), which is the part-whole relationship, and synonymy (Lin et al., 2003), which …

Content truncated to 500 words