In this project, we propose a Deep Learning architecture for Sarcasm Detection, while utilizing pre-trained Word Embeddings from three well-known models: Word2Vec, fastText and Glove.
By taking inspiration from CADE, our intuition is to use a non-trainable Embedding layer representing the three models and a Bidirectional Long Term Short Memory (Bi-LTSM) layer.
This model will also be compared to current state-of-the-art approaches for Sarcasm Detection in order to check if we can reach comparable results.
We also try to understand why it works from a semantic point of view. Does the accuracy of sarcasm detection depend on the nature of word embeddings?