metadata

language:
  - en
license: unknown
size_categories:
  - 10K<n<100K
task_categories:
  - feature-extraction
  - text-classification
  - image-classification
  - image-feature-extraction
  - zero-shot-classification
  - zero-shot-image-classification
pretty_name: multimodal-sarcasm-dataset
tags:
  - sarcasm
  - sarcasm-detection
  - mulitmodal-sarcasm-detection
  - sarcasm detection
  - multimodao sarcasm detection
  - tweets
dataset_info:
  - config_name: mmsd-original
    features:
      - name: image
        dtype: image
      - name: text
        dtype: string
      - name: label
        dtype: int64
    splits:
      - name: train
        num_bytes: 1816409874.384
        num_examples: 19816
      - name: validation
        num_bytes: 260024770
        num_examples: 2410
      - name: test
        num_bytes: 262626922.717
        num_examples: 2409
    download_size: 2690054686
    dataset_size: 2339061567.101
  - config_name: mmsd-v1
    features:
      - name: image
        dtype: image
      - name: text
        dtype: string
      - name: label
        dtype: int64
    splits:
      - name: train
        num_bytes: 1797521611.232
        num_examples: 19557
      - name: validation
        num_bytes: 259452303.817
        num_examples: 2387
      - name: test
        num_bytes: 261557636.749
        num_examples: 2373
    download_size: 2667548595
    dataset_size: 2318531551.798
  - config_name: mmsd-v2
    features:
      - name: image
        dtype: image
      - name: text
        dtype: string
      - name: label
        dtype: int64
    splits:
      - name: train
        num_bytes: 1816105257.384
        num_examples: 19816
      - name: validation
        num_bytes: 259989983
        num_examples: 2410
      - name: test
        num_bytes: 262588464.717
        num_examples: 2409
    download_size: 2689804711
    dataset_size: 2338683705.101
configs:
  - config_name: mmsd-original
    data_files:
      - split: train
        path: mmsd-original/train-*
      - split: validation
        path: mmsd-original/validation-*
      - split: test
        path: mmsd-original/test-*
  - config_name: mmsd-v1
    data_files:
      - split: train
        path: mmsd-v1/train-*
      - split: validation
        path: mmsd-v1/validation-*
      - split: test
        path: mmsd-v1/test-*
  - config_name: mmsd-v2
    data_files:
      - split: train
        path: mmsd-v2/train-*
      - split: validation
        path: mmsd-v2/validation-*
      - split: test
        path: mmsd-v2/test-*

MMSD2.0: Towards a Reliable Multi-modal Sarcasm Detection System

This is a copy of the dataset uploaded on Hugging Face for easy access. The original data comes from this work, which is an improvement upon a previous study.

Usage

from typing import TypedDict, cast

import pytorch_lightning as pl
from datasets import Dataset, load_dataset
from torch import Tensor
from torch.utils.data import DataLoader
from transformers import CLIPProcessor


class MMSDModelInput(TypedDict):
    pixel_values: Tensor
    input_ids: Tensor
    attention_mask: Tensor
    label: Tensor
    id: list[str]


class MMSDDatasetModule(pl.LightningDataModule):

    def __init__(
        self,
        clip_ckpt_name: str = "openai/clip-vit-base-patch32",
        dataset_version: str = "mmsd-v2",
        max_length: int = 77,
        train_batch_size: int = 32,
        val_batch_size: int = 32,
        test_batch_size: int = 32,
        num_workers: int = 19,
    ) -> None:
        super().__init__()
        self.clip_ckpt_name = clip_ckpt_name
        self.dataset_version = dataset_version
        self.train_batch_size = train_batch_size
        self.val_batch_size = val_batch_size
        self.test_batch_size = test_batch_size
        self.num_workers = num_workers
        self.max_length = max_length

    def setup(self, stage: str) -> None:
        processor = CLIPProcessor.from_pretrained(self.clip_ckpt_name)

        def preprocess(example):
            inputs = processor(
                text=example["text"],
                images=example["image"],
                return_tensors="pt",
                padding="max_length",
                truncation=True,
                max_length=self.max_length,
            )

            return {
                "pixel_values": inputs["pixel_values"],
                "input_ids": inputs["input_ids"],
                "attention_mask": inputs["attention_mask"],
                "label": example["label"],
            }

        self.raw_dataset = cast(
            Dataset,
            load_dataset("coderchen01/MMSD2.0", name=self.dataset_version),
        )
        self.dataset = self.raw_dataset.map(
            preprocess,
            batched=True,
        )

    def train_dataloader(self) -> DataLoader:
        return DataLoader(
            self.dataset["train"],
            batch_size=self.train_batch_size,
            shuffle=True,
            num_workers=self.num_workers,
        )

    def val_dataloader(self) -> DataLoader:
        return DataLoader(
            self.dataset["validation"],
            batch_size=self.val_batch_size,
            num_workers=self.num_workers,
        )

    def test_dataloader(self) -> DataLoader:
        return DataLoader(
            self.dataset["test"],
            batch_size=self.test_batch_size,
            num_workers=self.num_workers,
        )

References

[1] Yitao Cai, Huiyu Cai, and Xiaojun Wan. 2019. Multi-Modal Sarcasm Detection in Twitter with Hierarchical Fusion Model. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2506–2515, Florence, Italy. Association for Computational Linguistics.

[2] Libo Qin, Shijue Huang, Qiguang Chen, Chenran Cai, Yudi Zhang, Bin Liang, Wanxiang Che, and Ruifeng Xu. 2023. MMSD2.0: Towards a Reliable Multi-modal Sarcasm Detection System. In Findings of the Association for Computational Linguistics: ACL 2023, pages 10834–10845, Toronto, Canada. Association for Computational Linguistics.