Parrot Multilingual Eval

parrot-multilingual-eval

Parrot: Multilingual Visual Instruction Tuning — Sun et al. (2024) (arXiv:2406.02539, 2024)

What this evaluates

Evaluates the multilingual visual-language understanding capabilities of multimodal large language models (MLLMs) across six languages (English, Chinese, Portuguese, Arabic, Turkish, Russian). It probes how well models align visual features with non-English textual instructions and handle cross-lingual multimodal tasks without relying on naive translation.

Datasets

• MMMB — total ?; splits: test (-1)
• MMBench — total ?; splits: test (-1)

Metrics

• Accuracy (primary) — range: percent
  • Percentage of correctly answered questions out of the total number of questions. Calculated as (number of correct predictions / total questions) × 100.

Input / output format

Input: An image paired with a multilingual text prompt or question in one of six languages (English, Chinese, Portuguese, Arabic, Turkish, Russian).

Output: Text response (typically multiple-choice selection or descriptive answer) generated by the MLLM.

Scoring recipe

def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for pred, gold in zip(predictions, gold_labels) if pred == gold)
    return (correct / len(gold_labels)) * 100

Common pitfalls

• Relying on naive machine translation for non-English queries causes a 'seesaw effect', degrading performance in some languages while improving others.
• Multilingual datasets often suffer from translation noise and class imbalance, which can trigger the 'curse of multilingualism' if not handled carefully.
• Evaluating only on high-resource languages (English/Chinese) masks performance degradation in low-resource settings.

Evidence (verbatim from paper)

Table 1: Accuracy performance comparison on multilingual benchmarks. We report all compared methods with VLMEvalKit*(Duan et al., [2024]). The best and second results are shown in bold and underline, respectively. Our evaluation consists of two parts: one assessing the multilingual capabilities of MLLMs, while the other evaluating its overall performance. The first part is conducted on two datasets: multilingual MMBench(Liu et al., [2023c])* and a newly developed benchmark MMMB.

Citation

@misc{sun2024parrot,
  title={Parrot: Multilingual Visual Instruction Tuning},
  author={Sun et al. (2024)},
  year={2024},
  note={arXiv:2406.02539}
}

• arXiv: 2406.02539