DISCLAIMER: WORK IN PROGRESS >> WILL BE UPDATED WITH MORE INFORMATION
This model based on BERTweet-base has been finetuned on the Semeval 2018 Task 3 dataset for Irony Detection in English. However, we do not use the original labels, as we have provided more finegrained labels and annotated the tweets without irony-related hashtags. These specific models are based on a new paper accepted at the Joint LREC-COLING main conference. This model was trained on 4,592 samples (not the standard benchmark dataset) to be evaluated on 200 tweets labelled by three different annotators.
A model with the same parameters was also evaluated on the complete dataset through 10-fold CV. TODO: add the scores once the paper is published.
REFERENCE TO THE PAPER WILL BE INCLUDED ONCE IT IS PUBLISHED.