SEMANTIFY: Unveiling Memes with Robust Interpretability beyond Input Attribution

Authors: Dibyanayan Bandyopadhyay, Asmit Ganguly, Baban Gain, Asif Ekbal

IJCAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Evaluation of SEMANTIFY using interpretability metrics, including leakage-adjusted simulatability, demonstrates its superiority over various baselines by up to 2.5 points. Human evaluation of relatedness and exhaustiveness of extracted keywords further validates its effectiveness. Additionally, a qualitative analysis of extracted keywords serves as a case study, unveiling model error cases and their reasons.
Researcher Affiliation Academia 1Department of Computer Science and Engineering, Indian Institute of Technology, Patna, India 2School of AI and Data Science, Indian Institute of Technology, Jodhpur, India
Pseudocode Yes Algorithm 1 Retrieve explainable keywords with four step filtering
Open Source Code Yes Code and Supplementary Material available at: https://github.com/newcodevelop/semantify
Open Datasets Yes We use the Facebook Hateful Meme dataset [Kiela et al., 2021] for performing the experiments.
Dataset Splits Yes To ensure robust evaluation on simulatability, we conduct a 5-fold cross-validation for testing the surrogate models (Section 4.2) after running experiments for 3, 500 steps on the respective train set.
Hardware Specification Yes All experiments were conducted on a single Nvidia A100 80GB GPU.
Software Dependencies No The paper mentions software like PyTorch, Python, GPT-2, and Huggingface transformers but does not provide specific version numbers for any of them.
Experiment Setup Yes We employed the Adam optimizer [Kingma and Ba, 2017] with a learning rate of 0.005 for optimization.