ALB at SemEval-2018 Task 10: A System for Capturing Discriminative Attributes

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 963–967 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics ALB at SemEval-2018 Task 10: A System for Capturing Discriminative Attributes Bogdan Dumitru, Alina Maria Ciobanu, Liviu P. Dinu Faculty of Mathematics and Computer Science, University of Bucharest Human Language Technologies Research Center, University of Bucharest bogdan27182@gmail.com, alina.ciobanu@my.fmi.unibuc.ro, ldinu@fmi.unibuc.ro Abstract Semantic difference detection attempts to cap- ture whether a word is a discriminative at- tribute between two other words. For ex- ample, the discriminative feature red charac- terizes the first word from the (apple, ba- nana) pair, but not the second. Modeling semantic difference is essential for language understanding systems, as it provides use- ful information for identifying particular as- pects of word senses. This paper describes our system implementation (the ALB system of the NLP@Unibuc team) for the 10th task of the Semaval 2018 workshop, “Capturing Discriminative Attributes”. We propose a method for semantic difference detection that uses an SVM classifier with features based on co-occurrence counts and shallow semantic parsing, achieving 0.63 F1 score in the compe- tition. 1 Introduction and Related Work Semantic similarity detection is a well-studied re- search problem with numerous applications. State of the art models are extremely capable, deter- mining the degree of semantic similarity between words with high accuracy. However, looking the other way around, the se- mantic difference between words has received sig- nificantly less attention. As a result, one can argue that a semantic similarity model without the capa- bility of spotting the semantic difference as well is not a complete system, and may not prove very useful in practice. Semantic difference is a ternary relation (w1, w2, w3), where w1 and w2 are called concepts and w3 is called discriminative feature. The dis- criminative feature characterizes the first concept, w1, but not the second one, w2. If the discrimina- tive feature characterizes both w1 and w2 or none of them, then we do not have semantic difference. Semantic difference detection is a binary clas- sification task where given a triplet of words, a model needs to determine if a semantic difference is present or not. As emphasized by Krebs and Paperno (2016), this non-trivial task has numer- ous applications, such as automatized lexicogra- phy, conversational agents or machine translation. Most research on discriminative features is re- lated to computer vision (Farhadi et al., 2009; Rus- sakovsky and Fei-Fei, 2012), as these attributes proved to be very useful in interpreting visual data (Huang et al., 2016), being able to link visual fea- tures and semantic labels (Guo et al., 2015). A recent study on this topic belongs to Lazaridou et al. (2016), who proposed a method for iden- tifying discriminative attributes when given word pairs and their visual representations. In this paper, we describe a system for semantic difference detection that outputs a set of features for every triplet in the input data, based on prepro- cessed external resources (the English Wikipedia database). Further, these features are …

Content truncated to 500 words