View on GitHub

SIDATA

Intended-Sarcasm-Detection-In-English-and-Arabic-for-extremly-unbalanced-datasets

URL: https://github.com/rematchka/Intended-Sarcasm-Detection-In-English-and-Arabic-for-extremly-unbalanced-datasets

Description: This repo contains work carried out for SemEval 2022 Task 6: iSarcasmEval: Intended Sarcasm Detection In English and Arabic.

Methods

The repository includes the implementation of the system used in SemEval-2022 Task 6: Intended Sarcasm Detection in English and Arabic. The proposed approach involves the following steps:

  1. Voting Classifier: A classifier combining multiple BERT-based models (KimCNN) to improve performance on an extremely unbalanced dataset.
  2. Modified Loss Functions: The model trains BERT-Base models with customized loss functions designed to address issues associated with unbalanced datasets.
  3. Handcrafted Features: Features are handcrafted and combined with machine learning models like SVM to improve detection performance.

Additionally, the repository investigates the importance of different layers in BERT-Base models and identifies appropriate loss functions to optimize deep learning models for the unbalanced data scenario.

Results

The proposed model demonstrated competitive performance by utilizing a voting classifier and addressing the challenges posed by an unbalanced dataset through the use of modified loss functions and ensemble methods.

Dataset

The dataset used in this work is part of SemEval-2022 Task 6 and can be accessed via the following repository:

This dataset includes English and Arabic data for the task of detecting intended sarcasm in text. You can find the detailed instructions for dataset usage and format in the repository.