Sentimaithili Eval

sentimaithili-eval

SentiMaithili: A Benchmark Dataset for Sentiment and Reason Generation for the Low-Resource Maithili Language — Ranjan et al. (2025) (arXiv:2510.22160, 2025)

What this evaluates

Evaluates sentiment classification and justification generation capabilities for the low-resource Maithili language. It probes a model's ability to accurately predict sentence-level sentiment labels and generate culturally grounded, linguistically correct explanations in Maithili.

Datasets

• SentiMaithili — total 3221; splits: test (-1)

Metrics

• Precision — range: [0, 1]
  • Measures the proportion of correctly predicted positive instances among all predicted positives.
• Recall — range: [0, 1]
  • Measures the proportion of correctly predicted positive instances among all actual positives.
• F1-score (primary) — range: [0, 1]
  • Harmonic mean of precision and recall, balancing both aspects. Macro-averaging is used to give equal importance to all classes regardless of frequency.
• BLEU — range: [0, 1]
  • Evaluates precision-based n-gram overlap up to 4-grams between generated and reference text.
• ROUGE-1 — range: [0, 1]
  • Measures unigram (word-level) overlap, emphasizing recall.
• ROUGE-L — range: [0, 1]
  • Measures the longest common subsequence (LCS) between the generated and reference text.

Input / output format

Input: Sentence-level text in the Maithili language.

Output: For classification: a sentiment label. For generation: a Maithili-written justification/rationale conditioned on the input sentence and predicted label.

Scoring recipe

def compute_metrics(preds, golds, gen_preds, gen_refs):
    # Classification (Macro-averaged)
    f1_scores = []
    for label in unique_labels:
        tp = sum(1 for p, g in zip(preds, golds) if p == g == label)
        fp = sum(1 for p, g in zip(preds, golds) if p == label and g != label)
        fn = sum(1 for p, g in zip(preds, golds) if p != label and g == label)
        prec = tp / (tp + fp) if (tp + fp) > 0 else 0
        rec = tp / (tp + fn) if (tp + fn) > 0 else 0
        f1_scores.append(2 * prec * rec / (prec + rec) if (prec + rec) > 0 else 0)
    macro_f1 = sum(f1_scores) / len(f1_scores)
    # Generation
    bleu = compute_bleu_4g(gen_preds, gen_refs)
    rouge1 = compute_rouge_unigram(gen_preds, gen_refs)
    rouge_l = compute_rouge_lcs(gen_preds, gen_refs)
    return macro_f1, bleu, rouge1, rouge_l

Common pitfalls

• Using micro-averaging instead of macro-averaging for F1-score, which would skew results toward majority classes.
• Evaluating justification generation with only exact match or perplexity instead of the specified n-gram overlap metrics (BLEU/ROUGE).
• Ignoring the two-stage hierarchical task structure, where generation must be conditioned on both the input sentence and the predicted sentiment label.

Evidence (verbatim from paper)

Classification task: Precision: Measures the proportion of correctly predicted positive instances among all predicted positives. Recall: Measures the proportion of correctly predicted positive instances among all actual positives. F1-score: Harmonic mean of precision and recall, balancing both aspects. Macro-averaging is used to give equal importance to all classes regardless of frequency. Justification generation task: BLEU: Evaluates precision-based n-gram overlap up to 4-grams between generated and reference text. ROUGE-1: Measures unigram (word-level) overlap, emphasizing recall. ROUGE-L: Measures the longest common subsequence (LCS) between the generated and reference text.

Citation

@misc{ranjan2025sentimaithili,
  title={SentiMaithili: A Benchmark Dataset for Sentiment and Reason Generation for the Low-Resource Maithili Language},
  author={Ranjan et al. (2025)},
  year={2025},
  note={arXiv:2510.22160}
}

• arXiv: 2510.22160