IITK at SemEval-2024 Task 4: MULTILINGUAL DETECTION OF PERSUASION TECHNIQUES IN MEMES

The following repository contains the source code for our submission in the Task 4 of SemEval-2024. The link to the paper

Table of Contents

• Directory Structure
• Running
  • HypEmo-Task1 - Training and Inference
  • HypEmo-Task2 - Training and Inference
  • CDP - Training and Inference
• Our Models:
  • Drive Directory Structure
• Code Reference

Directory Structure

IITK-SemEval-2024-Task-4-Persuation-Techniques/
├── HypEmo-Task1
│   ├── Hypemo
│   │   ├── data
│   │   ├── exp
│   │   ├── label_tree
│   │   └── train_label_embedding.py
│   ├── Merged_test_for_HypEmo_Task1.csv
│   ├── Results-Creation-Translation-HypEmo-Task1.ipynb
├── Hypemo-Task2
│   ├── Hypemo
│   │   ├── data
│   │   ├── exp
│   │   ├── label_tree
│   │   └── train_label_embedding.py
│   ├── Merged_test_for_HypEmo_Task2b.csv
│   ├── Results-Creation-Translation-HypEmo-Task2b.ipynb
├── Fine-Grained-Emotion
│   ├── FineGrained-Emotion-Prediciton-Using-Definitions-new
│   │   ├── ckpt
│   │   ├── config
│   │   ├── data
│   │   └── run_def_nsp-mlm.py
│   ├── Merged_test_for_CDP_2a.csv
│   ├── Merged_test_for_CDP_Task1.csv
│   ├── Results-Creation-Translation-CDP-Task1.ipynb
│   └── Results-Creation-Translation-CDP-Task2a.ipynb
├── README.md
└── Data-Creation-Techniques
    ├── CLIP-embeddings.ipynb
    ├── Training-Creation-Translation-HypEmo-2b.ipynb
    ├── Training-Creation-Translation-CDP-Task1.ipynb
    ├── Training-Creation-Translation-CDP-Task2b.ipynb
    └── Training-Creation-Translation-HypEmo-Task2b.ipynb

• HypEmo-Task1/
  • For training using HypEmo only and generating predictions for task-1.
• HypEmo-Task2/
  • For training using HypEmo only and generating predictions for task-2 using CLIP embeddings.
• Fine-Grained-Emotion/
  • Consists of codes utilised for CDP in Task-1 and Task-2.
• Data-Creation-Techniques/
  • CLIP-embeddings.ipynb
    • To create CLIP embeddings for the training and testing dataset
  • Training-Creation-Translation-HypEmo-2b.ipynb
    • To create training and testing dataset for HypEmo Task 2a
  • Training-Creation-Translation-CDP-Task1.ipynb
    • To create training and testing dataset for CDP Task 1
  • Training-Creation-Translation-CDP-Task2b.ipynb
    • To create training and testing dataset for CDP Task 2b
  • Training-Creation-Translation-HypEmo-Task2b.ipynb
    • To create training and testing dataset for HypEmo Task 2a

Running

HypEmo-Task1 - Training and Inference

• Using python train.py to generate the softmax predictions.
• Generate the predictions using Results-Creation-Translation-HypEmo-Task1.ipynb for task1.

HypEmo-Task2 - Training and Inference

• Using python train.py to generate the softmax predictions.
• Generate the predictions using Results-Creation-Translation-HypEmo-Task2b.ipynb for task 2a.

CDP - Training and Inference

• Using python run_def_nsp-mlm.py --taxonomy subtask_nsp-mlm_0.3.json to generate the predictions for task1.
• Using python run_def_nsp-mlm.py --taxonomy subtask_nsp-mlm_0.4.json to generate the predictions for task2a.
• Generate the predictions using Results-Creation-Translation-CDP-Task1.ipynb for task1.
• Generate the predictions using Results-Creation-Translation-CDP-Task2a.ipynb for task2a.

Our Models:

The models trained and used for submission can be found below: https://1drv.ms/f/s!AuBOJ2hW9GimhZsGvBc6Y_JrXf5Xtg?e=Nwtyar

Drive Directory Structure

Drive/
├── Models
│   ├── HypEmo-Task1
│   ├── HypEmo-Task2
│   └── Fine-Grained-Emotion
└── Submitted_files
    ├── Sub-Task 1
    ├── Sub-Task 2

• Models/

  • Consists of the trained models.

• Submitted-files/

  • Consists of the submitted files.

Code Reference

https://github.com/dinobby/HypEmo

https://github.com/Exploration-Lab/FineGrained-Emotion-Prediciton-Using-Definitions