Description:
The development and utilization of datasets in the context of LRLs remain a significant challenge. As there is no standard Nepali news headline dataset, we initially considered collecting data from a range of news portals. However, a comprehensive review of the literature on Nepali NLP tasks revealed the existence of the NepBERTa (https://nepberta.github.io/) , a BERT-based model trained on a large corpus comprising 0.8 billion words from 36 distinct popular news sites in Nepal, with a total of 2,762,486 news items. The publicly accessible dataset contains 2,760,921 news items covering diverse topics and categories. The dataset comprises multiple fields such as news categories, headings, text, dates, and links. In this study, only the news body (text) and corresponding headlines were selected from the dataset for further analysis. Several preprocessing steps were applied to ensure the quality and consistency of the dataset. First, 1,228 news items identified as non-Nepali were removed. Additionally, the dataset contains 148,674 news items with text lengths shorter than 50 words which were excluded to avoid insufficient context for headline generation. The dataset also includes instances with XML tags in both the body and headline sections. These tags were removed to ensure clean textual input. After preprocessing, the total number of news items with headlines was 2,611,019. Due to computational constraints and the large size of the original corpus, we developed a customized dataset, which we refer to as the NEPHEAD dataset. The NEPHEAD dataset contains the news body and headings categories from a random selection of 10% of the preprocessed data, resulting in 261,101 news-headline pairs. The NEPHEAD dataset is divided into training, validation, and test sets following an approximate 80:10:10 split. Training: 208,880 Validation: 26,111 Testing: 26,110
Linguistic Characteristics
Nepali uses Devanagari script, which introduces subword fragmentation challenges in multilingual Transformer tokenizers.
Intended Uses
- Headline generation
- Low-resource NLP research
- Seq2seq model evaluation
- Transformer architecture comparison
Limitations
- News-domain only
- Possible media-source bias
Benchmark Results
Automatic Evaluation Scores
| Model | ROUGE-1 (%) | ROUGE-2 (%) | ROUGE-L (%) | BLEU (%) | METEOR (%) | SBERT | BERTScore-F1 (%) |
|---|---|---|---|---|---|---|---|
| mT5-small (Conventional) | 7.68 | 2.25 | 7.42 | 1.80 | 5.94 | 40.72 | 84.84 |
| mT5-small (Fine-tuned) | 36.50 | 19.11 | 35.16 | 13.18 | 30.47 | 70.45 | 90.36 |
| mBART-large-cc25 (Conventional) | 21.76 | 9.50 | 20.42 | 5.53 | 24.51 | 62.57 | 87.14 |
| mBART-large-cc25 (Fine-tuned) | 37.67 | 20.38 | 36.06 | 14.20 | 34.46 | 71.62 | 90.58 |
| NepBERTa (Conventional) | 0.08 | 0.00 | 0.08 | 0.01 | 0.03 | 17.10 | 79.19 |
| NepBERTa (Fine-tuned) | 12.12 | 2.32 | 11.57 | 2.10 | 9.45 | 52.52 | 84.71 |
| LLaMA-3.2-1B (Fine-tuned) | 34.89 | 18.33 | 33.37 | 12.54 | 30.37 | 69.64 | 90.04 |
Human Evaluation Scores (1–5 Scale)
| Model | Relevance | Conciseness | Fluency | Accuracy | Engagement | Overall |
|---|---|---|---|---|---|---|
| mT5-small (Fine-tuned) | 4.22 | 4.14 | 4.23 | 4.06 | 4.10 | 4.15 |
| mBART-large-cc25 (Fine-tuned) | 4.36 | 4.23 | 4.40 | 4.32 | 4.29 | 4.32 |
| NepBERTa (Fine-tuned) | 1.95 | 1.64 | 1.44 | 1.55 | 1.54 | 1.62 |
| LLaMA-3.2-1B (Fine-tuned) | 4.09 | 3.89 | 3.94 | 3.82 | 3.66 | 3.88 |
Human Evaluation Agreement Analysis
| Metric | Avg Spearman (ρ) | Avg Pearson (r) |
|---|---|---|
| Relevance | 0.603 | 0.648 |
| Conciseness | 0.656 | 0.701 |
| Fluency | 0.688 | 0.745 |
| Accuracy | 0.656 | 0.709 |
| Engagement | 0.668 | 0.714 |
| Overall | 0.679 | 0.754 |
Standard Deviation of Human Evaluation Scores
| Model | Relevance (SD) | Conciseness (SD) | Fluency (SD) | Accuracy (SD) | Engagement (SD) |
|---|---|---|---|---|---|
| mT5-small | 0.86 | 0.94 | 0.96 | 1.03 | 0.93 |
| mBART-large-cc25 | 0.72 | 0.80 | 0.77 | 0.78 | 0.76 |
| NepBERTa | 1.11 | 0.88 | 0.78 | 0.82 | 0.87 |
| LLaMA-3.2-1B | 1.00 | 1.02 | 1.05 | 1.04 | 1.01 |
Citation
The paper has been published in ACM TALLIP.
Sunil Dahal and Katsuhide Fujita. 2026. Evaluation of Monolingual and Multilingual Transformer Models for Nepali Headline Generation. ACM Trans. Asian Low-Resour. Lang. Inf. Process. Just Accepted (May 2026). https://doi.org/10.1145/3816146
Ethical Considerations
The dataset was collected from publicly available Nepali news sources for research purposes only. No personally identifiable information was intentionally collected.
- Downloads last month
- 20