Wolves at SemEval-2018 Task 10: Semantic Discrimination based on Knowledge and Association

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 972–976 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics Wolves at SemEval-2018 Task 10: Semantic Discrimination based on Knowledge and Association Shiva Taslimipoor1, Omid Rohanian1, Le An Ha1, Gloria Corpas Pastor2 and Ruslan Mitkov1 1Research Group in Computational Linguistics, University of Wolverhampton, UK 2University of Malaga, Spain {shiva.taslimi, m.rohanian, l.a.ha, r.mitkov}@wlv.ac.uk gcorpas@uma.es Abstract This paper describes the system submitted to SemEval 2018 shared task 10 ‘Capturing Dis- criminative Attributes’. We use a combina- tion of knowledge-based and co-occurrence features to capture the semantic difference be- tween two words in relation to an attribute. We define scores based on association measures, ngram counts, word similarity, and Concept- Net relations. The system is ranked 4th (joint) on the official leaderboard of the task. 1 Introduction When it comes to investigating semantic similari- ties, it is worth noting that similarity between two words can be too general to quantify. Accordingly, the discriminating power of a model is also im- portant in limiting the scope of similarity between words. The main idea behind distributional semantics, known as Distributional Hypothesis (DH), states that linguistic items with similar distributions have similar meanings (Blevins, 2016). Therefore these methods are biased towards finding similarities between concepts. The SemEval shared task 10 ‘Capturing Discriminative Attributes’ poses the new problem of semantic difference detection, thus putting difference, rather than similarity at the forefront. It is about modeling semantic difference in the case of already related words. The idea is that while similarity can group words together in a generic way, understanding semantic differences sheds additional light on the meaning of each in- dividual word. A semantic model can potentially become more robust if it can benefit from sensitivity to differ- ences alongside similarities in meaning. Consid- ering difference can also help researchers assess semantic representations more rigorously. The effectiveness of a semantic similarity model can be evaluated further by quantifying its strength in finding differences between words. In the shared task, semantic difference is oper- ationalised as the relation between two semanti- cally related words and a discriminative feature. This relation is realised if the feature characterises only the first word. An example is the triple air- plane, helicopter, wings. In this formulation, se- mantic difference is an asymmetric relation. In this work, we compute several scores for word pairs and triples with the aim of capturing different semantic relations. Specifically, we de- fine scores based on a knowledge-based ontology and co-occurrence counts. For knowledge-based features we rely on ConceptNet semantic network (Speer and Havasi, 2013), and our co-occurrence based features are derived from association mea- sures, ngrams and pre-trained embeddings. We use the scores in both supervised and unsupervised scenarios to identify triples that constitute seman- tic difference; i.e. the attribute (third) word is dis- criminative between the first two words. The code and data used for this system are freely available.1 The rest of this paper is organised as follows: Section 2 describes related …

Content truncated to 500 words