Meaning_space at SemEval-2018 Task 10: Combining explicitly encoded knowledge with information extracted from word embeddings

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 940–946 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics Meaning space at SemEval-2018 Task 10: Combining explicitly encoded knowledge with information extracted from word embeddings Pia Sommerauer Antske Fokkens Piek Vossen Computational Lexicology & Terminology Lab (CLTL) Vrije Universiteit Amsterdam, the Netherlands {pia.sommerauer,antske.fokkens,piek.vossen}@vu.nl Abstract This paper presents the two systems submitted by the meaning space team in Task 10 of the SemEval competition 2018 entitled Capturing discriminative attributes. The systems consist of combinations of approaches exploiting ex- plicitly encoded knowledge about concepts in WordNet and information encoded in distri- butional semantic vectors. Rather than aim- ing for high performance, we explore which kind of semantic knowledge is best captured by different methods. The results indicate that WordNet glosses on different levels of the hier- archy capture many attributes relevant for this task. In combination with exploiting word em- bedding similarities, this source of information yielded our best results. Our best performing system ranked 5th out of 13 final ranks. Our analysis yields insights into the different kinds of attributes represented by different sources of knowledge. 1 Introduction SemEval Task 10 “Capturing Discriminative At- tributes” (Krebs et al., 2018) provides participants with triples of words consisting of two concepts and an attribute. The task is to determine whether the attribute is a distinguishing property of the first concept compared to the second concept. This is the case in triple shrimp, spinach, pink, for instance, because shrimp can be pink whereas spinach is usually of a different color. When the first concept does not have a semantic relation with the attribute or both the concepts have the same se- mantic relation with it, the attribute is considered not to be discriminative. In general, Task 10 can be understood as de- tecting whether there is a semantic relation be- tween the concepts and the attribute. The dataset includes a wide range of variation. For instance, the attribute may be a part of the concept (e.g. tortoise, snail, legs) or category membership (e.g. polyurethane, polyester, material), relations be- tween entities and activities they engage in (e.g. cheetah, lion, runs) as well as rather specific rela- tions, for instance the relation between a specialist and the phenomenon they are specialized in (e.g. optician, dentist, eyes). Rather than finding specific solutions for each kind of relation, we investigate different ap- proaches exploiting different sources of knowl- edge. Both of our systems comprise a component exploiting the glosses and hierarchical structure of WordNet (Fellbaum, 1998) in order to determine whether an attribute applies to a concept. Our un- derlying assumption is that definitions should pro- vide the most important distinctive attributes of a concept. Since concepts are not necessarily al- ways distinguished on the same level of concrete- ness, but might also be distinguished on a more ab- stract level (e.g. herbs, root, green v.s. apse, night- gown, royal) we exploit the entire WordNet hier- archy. In both of our systems, the second …

Content truncated to 500 words