Tel(s)-Telle(s)-Signs: Highly Accurate Automatic Crosslingual Hypernym Discovery

Tel(s)-Telle(s)-Signs: Highly Accurate Automatic Crosslingual Hypernym Discovery Ada Wan University of Zurich Zurich, Switzerland ada.wan@uzh.ch Abstract We report a highly accurate hypernym discovery heuristic that works on unrestricted texts. This approach leverages morphological cues in French, but given any parallel data and word alignment tool, this proves to be a technique that can work reliably in other languages as well. We tested this method using two French-English corpora of different genres (medical and news) and attained near-perfect accuracy. The key idea is to exploit morphological information in the French trigger phrase tel(s)-/telle(s)- que (meaning “such as” in English) to uniquely identify the correct hypernym. This shows to be an inexpensive and effective heuristic also when there are multiple noun phrases preceding the trigger phrase, as in the case of prepositional phrase attachment causing ambiguity in interpretation and hypernym acquisition, indicating that this pattern in French is more informative than its English counterpart. Keywords: hypernym discovery, information extraction, crosslingual knowledge transfer 1. Introduction The present work focuses on discovering hypernyms in French (FR) using the trigger phrase tel-1 que (meaning “such as” in English (EN)), exploiting morphological in- formation to uniquely identify the correct hypernym. This shows to be an inexpensive and effective heuristic also when there are multiple noun phrases preceding the trigger phrase, as in the case of prepositional phrase attachment causing ambiguity in interpretation and hypernym acqui- sition, proving that this pattern in FR is more informative than its EN counterpart. In simpler cases, where there is only one noun phrase (NP) preceding the trigger phrase such as, hypernym discovery is straightforward. For instance, in the sentence: [1] Les agrumes tels que l’orange, le citron ou le pample- mousse contiennent beaucoup de vitamin C.2 and its EN translation: “Citrus fruits such as oranges, lemons or grapefruits contain a lot of vitamin C.” the NPs preceding such as and tels que – (cit- rus) fruits and agrumes, respectively – are the hy- pernyms denoting the categories or classes in which the hyponyms oranges/orange, lemons/citron, and grape- fruits/pampelmousse belong. But seldom does one notice that the bound morpheme -s in agrumes and tels in FR also reveals agreement between the two words and has a role to play in hypernym recognition. It appears that one could use this morphological informa- tion in tel- to correctly identify the antecedent of this hyper- /hyponym relation, disambiguating cases where multiple nouns preceding tel- que (hereafter: tq) pose as plausi- ble hypernym candidates. This would also resolve the is- sue brought about by the prepositional compound phrase 1tel- is used as a shorthand for tel, tels, telle, or telles, cf. Sec- tion 3.1. 2http://bdl.oqlf.gouv.qc.ca/bdl/gabarit\ _bdl.asp?id=2437 construction, i.e. the combination of a noun phrase and a prepositional phrase (Lefever et al., 2014), alternatively formulated as the error caused by the “local nature of the patterns” (Ritter et al., 2009): “A sentence with ... urban birds in cities such as pigeons ... matches the pattern ‘C such as E’ with C bound to city and E …

Content truncated to 500 words