CitiusNLP at SemEval-2018 Task 10: The Use of Transparent Distributional Models and Salient Contexts to Discriminate Word Attributes

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 953–957 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics CitiusNLP at SemEval-2018 Task 10: The Use of Transparent Distributional Models and Salient Contexts to Discriminate Word Attributes Pablo Gamallo Centro Singular de Investigaci´on en Tecnolox´ıas da Informaci´on (CiTIUS) Universidade de Santiago de Compostela, Galiza pablo.gamallo@usc.es Abstract This article describes the unsupervised strat- egy submitted by the CitiusNLP team to Se- mEval 2018 Task 10, a task which consists of predicting whether a word is a discrimina- tive attribute between two other words. The proposed strategy relies on the correspondence between discriminative attributes and relevant contexts of a word. More precisely, the method uses transparent distributional models to extract salient contexts of words which are identified as discriminative attributes. The sys- tem performance reaches about 70% accuracy when it is applied on the development dataset, but its accuracy goes down (63%) on the offi- cial test dataset. 1 Introduction The goal of SemEval-2018 Task 10 (Paperno, Lenci and Krebs, To Appear) is to predict whether a word is a discriminative attribute between two other words. The key idea underlying this task is to capture semantic attributes of words in or- der to discriminate their senses. Distributional se- mantics is based on the assumption that two words have similar senses if they tend to appear with the same contextual words (Firth, 1957). As con- textual words actually refer to the semantic at- tributes of a given word, I will focus on identifying the most salient word contexts. So, my method to identify discriminative attributes relies on the identification of salient contexts, since they repre- sent the main semantic attributes of a word. For this purpose, in this paper we will make use of distributional models built with transparent and lexico-syntactic contexts. To capture discrimina- tive attributes, I will rank the most relevant con- texts of a word by using lexical association mea- sures between a given word and their contexts. My method is unsupervised and only requires pre- trained distributional models. This paper is organized as follows. The method is described in Section 2. Experiments, results, and a discussion on them are presented in Section 3. Finally, conclusions are addressed in Section 4. 2 The method As mentioned in the previous section, discrimi- native attributes might be captured by searching for the most salient contexts of words. For this purpose, the distributional vector space I have adopted is a transparent count-based model with explicit and sparse dimensions. Sparseness re- duction is performed by selecting the most salient contexts per word using a filtering strategy (Bor- dag, 2008; Gamallo and Bordag, 2011; Gamallo, 2017). The filtering strategy to select the most salient contexts consists of selecting, for each word, the S (salient) contexts with highest lexi- cal association scores (e.g. loglikelihood, ppmi, etc). The top S contexts are considered to be the most relevant and informative for each word. S is a global, arbitrarily defined constant whose usual values …

Content truncated to 500 words