Every Child Should Have Parents: A Taxonomy Refinement Algorithm Based on Hyperbolic Term Embeddings

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4811–4817 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 4811 Every child should have parents: a taxonomy refinement algorithm based on hyperbolic term embeddings Rami Aly1, Shantanu Acharya2, Alexander Ossa1, Arne K¨ohn4,1, Chris Biemann1, and Alexander Panchenko3,1 1Universit¨at Hamburg, Hamburg, Germany 2National Institute of Technology Mizoram, Aizawl, India 3Skolkovo Institute of Science and Technology, Moscow, Russia 4Saarland University, Saarbr¨ucken, Germany {5aly,2ossa,koehn,biemann,panchenko}@informatik.uni-hamburg.de Abstract We introduce the use of Poincar´e embeddings to improve existing state-of-the-art approaches to domain-specific taxonomy induction from text as a signal for both relocating wrong hy- ponym terms within a (pre-induced) taxonomy as well as for attaching disconnected terms in a taxonomy. This method substantially improves previous state-of-the-art results on the SemEval-2016 Task 13 on taxonomy ex- traction. We demonstrate the superiority of Poincar´e embeddings over distributional se- mantic representations, supporting the hypoth- esis that they can better capture hierarchi- cal lexical-semantic relationships than embed- dings in the Euclidean space. 1 Introduction The task of taxonomy induction aims at cre- ating a semantic hierarchy of entities by using hyponym-hypernym relations – called taxonomy – from text corpora. Compared to many other do- mains of natural language processing that make use of pre-trained dense representations, state-of- the-art taxonomy learning is still highly relying on traditional approaches like extraction of lexical- syntactic patterns (Hearst, 1992) or co-occurrence information (Grefenstette, 2015). Despite the suc- cess of pattern-based approaches, most taxonomy induction systems suffer from a significant number of disconnected terms, since the extracted relation- ships are too specific to cover most words (Wang et al., 2017; Bordea et al., 2016). The use of dis- tributional semantics for hypernym identification and relation representation has thus received in- creasing attention (Shwartz et al., 2016). How- ever, Levy et al. (2015) observe that many pro- posed supervised approaches instead learn proto- typical hypernyms (that are hypernyms to many other terms), not taking into account the rela- tion between both terms in classification. There- fore, past applications of distributional seman- tics appear to be rather unsuitable to be directly applied to taxonomy induction as the sole sig- nal (Tan et al., 2015; Pocostales, 2016). We ad- dress that issue by introducing a series of sim- ple and parameter-free refinement steps that em- ploy word embeddings in order to improve ex- isting domain-specific taxonomies, induced from text using traditional approaches in an unsuper- vised fashion. We compare two types of dense vector em- beddings: the standard word2vec CBOW model (Mikolov et al., 2013a,b), that embeds terms in Euclidean space based on distributional similarity, and the more recent Poincar´e embeddings (Nickel and Kiela, 2017), which capture similarity as well as hierarchical relationships in a hyperbolic space. The source code has been published1 to recreate the employed embedding, to refine taxonomies as well as to enable further research of Poincar´e em- beddings for other semantic tasks. 2 Related Work The extraction of taxonomic relationships from text corpora is a long-standing …

Content truncated to 500 words