Contributions to the Computational Treatment of Non-literal Language

Abstract

Non-literal language concerns the deliberate use of language in such a way that meaning cannot be inferred through a mere literal interpretation. In this thesis, three different forms of this phenomenon are studied; namely, irony, non-compositional Multiword Expressions (MWEs), and metaphor. We start by developing models to identify ironic comments in the context of the social micro-blogging website Twitter. In these experiments, we proposed a new way to extract features based on a study of their spatial structure. The proposed model is shown to perform competitively on a standard Twitter dataset. Next, we extensively study MWEs, which are the central point of focus in this work. We start by framing the task of MWE identi fication as sequence labelling and devise experiments to see the effect of eye-tracking data in capturing formulaic MWEs using structured prediction. We also develop a novel neural architecture to speci fically address the issue of discontinuous MWEs using a combination of Graph Convolutional Neural Networks (GCNs) and self-attention. The proposed model is subsequently tested on several languages where it is shown to outperform the state-of-the-art in overall criteria and also in capturing gappy MWEs. In the final part of the thesis, we look at metaphor and its interaction with verbal MWEs. In a series of experiments, we propose a hybrid BERT-based model augmented with a novel variation of GCN where we perform classifi cation on two standard metaphor datasets using information from MWEs. This model which performs at the same level with state-of-the-art is, to the best of our knowledge, the first MWE-aware metaphor identifi cation system paving the way for further experimentation on the interaction of different types of fi gurative language.