Memotion 2.0 Eval

---
name: memotion-2.0-eval
description: Evaluates models on classifying social media memes for sentiment, emotion intensity (humour, sarcasm, offensiveness), and motivation. It probes the ability of text-only and multi-modal architectures to perform both binary and ordinal/multi-class classification across multiple related subtasks. Use when the user wants to benchmark on Memotion 2.0, or asks about evaluating this task. Reports weighted F1.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2202.07543
  bibtex_key: bucur2022blue
  confidence: high
---

# memotion-2.0-eval

> BLUE at Memotion 2.0 2022: You have my Image, my Text and my Transformer — Bucur et al. (2022) (arXiv:2202.07543, 2022)

## What this evaluates

Evaluates models on classifying social media memes for sentiment, emotion intensity (humour, sarcasm, offensiveness), and motivation. It probes the ability of text-only and multi-modal architectures to perform both binary and ordinal/multi-class classification across multiple related subtasks.

## Datasets

- **Memotion 2.0** — total ?; splits: train (-1), val (-1), test (-1)

## Metrics

- `weighted F1` **(primary)** — range: [0, 1]
  - The weighted average of per-class F1 scores, where each class's F1 is multiplied by its support (number of true instances) and divided by the total number of instances.

## Input / output format

**Input**: Meme images and associated text captions, optionally processed into embeddings (e.g., CLIP features, image features) and text tokens.

**Output**: Predicted class labels for each emotion subtask (Sentiment, Humour, Sarcasm, Offensive, Motivation) in either binary or ordinal/multi-class format depending on the task definition.

## Scoring recipe

```python
def compute_weighted_f1(y_true, y_pred, classes):
    total_weight = 0
    weighted_f1_sum = 0
    for c in classes:
        tp = sum(1 for t, p in zip(y_true, y_pred) if t == c and p == c)
        fp = sum(1 for t, p in zip(y_true, y_pred) if t != c and p == c)
        fn = sum(1 for t, p in zip(y_true, y_pred) if t == c and p != c)
        precision = tp / (tp + fp) if (tp + fp) > 0 else 0
        recall = tp / (tp + fn) if (tp + fn) > 0 else 0
        f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0
        support = sum(1 for t in y_true if t == c)
        weighted_f1_sum += f1 * support
        total_weight += support
    return weighted_f1_sum / total_weight if total_weight > 0 else 0
```

## Common pitfalls

- Task B uses binary classification while Task C uses ordinal/multi-class intensity scales; models must map non-zero intensity predictions to the positive class for Task B if trained on Task C.
- Multi-task learning benefits vary by emotion; joint training improves humour, sarcasm, and offensiveness but may slightly hurt sentiment classification compared to single-task models.
- Modality ablation shows small differences; adding CLIP features improves results marginally, but the best submission combines all available modalities.

## Evidence (verbatim from paper)

> In order to measure the benefits of multi-task learning for classifying emotion intensities, we performed an ablation study by comparing the weighted F1 scores computed for each emotion intensity predictions made by models trained in two settings.

## Citation

```bibtex
@misc{bucur2022blue,
  title={BLUE at Memotion 2.0 2022: You have my Image, my Text and my Transformer},
  author={Bucur et al. (2022)},
  year={2022},
  note={arXiv:2202.07543}
}
```

- arXiv: 2202.07543