gretagandolfi / antonyms

CL project

Antonyms

Computational Linguistics course project (MSc Cognitive Science - Language and Multimodal Interaction Track, CIMeC @ University of Trento)

Abstract

In daily conversations, the negation of a predicate implies the truth of its nearest alternatives. This hypothesis is the core of the conversational negation, an exciting proposal by Kruszewski et al. (2016) that manages to combine this semantic theory of negation based on alternativehood with data from distributional semantics. The proposal works amazingly - both in a theoretical and a computational sense - with nouns. In this work, I tried to apply this proposal to pairs of adjectives with opposite meaning. More precisely, I tried to use this concept of negation and alternativehood in order to test if the purely theoretical difference between binary and non-binary antonyms is intuitive and if it can be derived from distributional data. Due to this fact, I worked on two different levels: the level of intuitions and the level of distributional semantics. First, I tested the intuitions of 26 non-native English speakers, and I tried to see if every adjective, when negated, elicit in speakers’ mind the same pattern of alternatives. My project relies on the idea that the number of alternatives suggested by the negation of an adjective can be a factor to use in order to classify the relation between an adjective and its antonym. The simple intuition at the core of this reasoning is: fewer alternatives a negated adjective suggests, more its relation with its antonym is binary. In this context, I tried to test if other speakers share my intuition. Secondly, I checked if the semantic difference between graded and non-graded (non- binary and binary) antonyms can be seen also from a distributional point of view, using the analogy task. The aim of my project was to find, or - at least - to suggest empirical proves of the theoretic distinction between graded and non-graded antonyms. Unfortunately, both tests happened to be useless to detect this kind of semantic distinction.

The repository contains:

• Jupiter notebook (selection of antonyms)
• report (pdf)