Nice Perfume. How Long Did You Marinate in It? Multimodal Sarcasm Explanation

Authors: Poorav Desai, Tanmoy Chakraborty, Md Shad Akhtar10563-10571

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results demonstrate convincing results over various baselines (adopted for Mu SE) across five evaluation metrics. We also conduct human evaluation on predictions and obtain Fleiss Kappa score of 0.4 as a fair agreement among 25 evaluators.
Researcher Affiliation Academia Poorav Desai, Tanmoy Chakraborty, Md Shad Akhtar Indraprastha Institute of Information Technology, Delhi (IIIT Delhi), India {desai19010@iiitd.ac.in, tanmoy, shad.akhtar}@iiitd.ac.in
Pseudocode No The paper includes architectural diagrams (Figure 3) but no formal pseudocode or algorithm blocks.
Open Source Code Yes Reproducibility: The source code and dataset are available at https://github.com/LCS2-IIITD/Multimodal Sarcasm-Explanation-Mu SE.
Open Datasets Yes To address Mu SE, we curate MORE, a novel multimodal sarcasm explanation dataset, consisting of 3510 sarcastic posts with natural language explanations manually generated by expert annotators.
Dataset Splits Yes We perform experiments on MORE and use 85:5:10 split to create train (2983), validation (175), and test (352) sets.
Hardware Specification No The paper mentions training details like 'batch size = 16' and '125 epochs' but does not specify any particular GPU models, CPU models, or other hardware specifications.
Software Dependencies No The paper mentions 'BART (Lewis et al. 2020) tokenizer' and 'Adam W (Loshchilov and Hutter 2017) optimizer' but does not provide specific version numbers for these or other software libraries.
Experiment Setup Yes Hyperparameters: We employ BART (Lewis et al. 2020) tokenizer with maximum token length as 256. We use Adam W (Loshchilov and Hutter 2017) optimizer with learning rate of 1e 5 for the single cross-modal encoder and 3e 4 for the LM head of decoder. We train Ex More for 125 epochs with batch size = 16. During training, the cross-entropy loss is monitored over the validation set with image encoder in a frozen state.