Figurative Language Understanding based on Large Language Models

In the vast realm of Natural Language Processing, one of the areas that still presents a bottleneck is Figurative Language Understanding. However, this field is of fundamental interest both theoretically and practically; in all new applications of Artificial Intelligence, there is an increasing demand for a correct understanding of human language, which is naturally rich in rhetorical figures. To comprehend this figurative language, it is necessary for the model used to grasp all the different nuances and reasons, going beyond the mere literal meaning. Until now, attempts to solve this problem have primarily involved training specific models on large databases of rhetorical figures. In this thesis, we seek to overcome this challenge by utilizing a Large Language Model, specifically LLaMA. These kinds of models have already demonstrated immense potential in many Natural Language Processing tasks, and with this project, we have observed that it is possible to achieve promising performance simply by harnessing the potential of Large Language Models without specifically training them for our specific goal. In particular, we explore the use of LLaMA employing different types of prompting and through two separate tasks. The first one involves entailment classification of figurative language, while the second one is a text generation task where the model is asked to generate an explanation for the previously made choice. Through this exploration, we aim to contribute to the understanding of figurative language, emphasizing the transformative potential of Large Language Models in deciphering figurative expressions.