Gauurab/Nepali_Sentiment

NEPSEE: A Double-Annotated Benchmark for Nepali Sentiment Analysis

Please cite our preprint if you use this data. If this dataset, or anything derived from it, is used in a paper, model, or other work, please cite NEPSEE (arXiv:2610.19190). The BibTeX entry is in Citation.

NEPSEE is a benchmark of 3,189 Nepali sentences from public tweets, labelled for sentiment. Two annotators labelled each sentence independently and without seeing each other's labels. The benchmark keeps only sentences where they agreed on polarity.

• Paper: NEPSEE: A Double-Annotated Benchmark for Nepali Sentiment Analysis (Baral, Nepal, and Basnet, 2026)
• Language: Nepali (ne), Devanagari script
• Tasks: 3-class polarity and 5-point sentiment classification
• Date range of tweets: December 2017 to August 2026

Quick start

All rows are in one file, nepsee_sentiment.csv, and the benchmark split is stored in the split column. The Hugging Face viewer and load_dataset show the whole file as a single split named train, so filter on the column to get the real splits:

from datasets import load_dataset

ds = load_dataset("Gauurab/Nepali_Sentiment", split="train")  # all 3,189 rows

train = ds.filter(lambda row: row["split"] == "train")
dev = ds.filter(lambda row: row["split"] == "dev")
test = ds.filter(lambda row: row["split"] == "test")

Data fields

Column	Description
item_id	Unique sentence ID, in the form tweet_id:sentence_index
tweet_id	ID of the source tweet
sentence_index	Position of the sentence within its tweet, starting at 0
handle	X (Twitter) handle of the public account that posted the tweet
created_at	Tweet timestamp
sentence_text	The Nepali sentence
label_a	Five-point label from annotator A
label_b	Five-point label from annotator B
label3	Three-class gold label: the polarity both annotators agreed on (negative, neutral, positive)
label5	Five-point gold label. If the annotators differ only in intensity, this is the milder label
label5_alt	The stronger label if the annotators differ only in intensity, otherwise the same as label5
split	train, dev, or test

The five-point scale is strongly_negative, negative, neutral, positive, strongly_positive.

Using the five-point labels

On 2,423 sentences the two annotators chose exactly the same five-point label, so label5 equals label5_alt. On the other 766 they agreed on polarity but differed in intensity, for example positive and strongly_positive. Both labels are acceptable for those sentences:

• Lenient evaluation: count a prediction as correct if it matches either label5 or label5_alt.
• Strict evaluation: score only the sentences where label5 equals label5_alt.

label5 alone leans toward milder classes, so it should not be treated as the only correct answer on those 766 sentences.

Splits and label distribution

The split is a stratified 70/15/15 split grouped by tweet, so sentences from the same tweet never appear in different splits.

Split	Sentences	Negative	Neutral	Positive
train	2,235	443	710	1,082
dev	475	81	161	233
test	479	84	167	228
Total	3,189	608	1,038	1,543

How the data was created

Collection. Tweets were collected from the timelines of public figures in Nepali politics, journalism, and civil society. The sentences in this benchmark come from 2,381 tweets by 11 accounts.

Text processing. Tweets with no Devanagari text were removed, and Latin-script text, including English words, URLs, and Latin hashtags, was stripped from the rest. Tweets were then split into sentences on the danda (।) and on newlines, and sentences shorter than five words were dropped.

Annotation and selection. Two annotators labelled sentences independently and blind to each other, using a five-point scale plus a dont_know option.

Step	Sentences
Labelled by both annotators	5,614
Removed: either annotator chose dont_know	671
Removed: annotators disagreed on polarity	1,704
Removed: duplicates	50
Kept in the benchmark	3,189

Agreement. On the 4,943 sentences that both annotators gave a sentiment label, before filtering:

Measure	Value
Same polarity	65.5%
Cohen's kappa, 3-class polarity	0.465
Cohen's kappa, 5-point	0.289
Krippendorff's alpha, 5-point ordinal	0.625
Within one step on the 5-point scale	95.6%

Limitations

• Filtered for agreement. Only sentences with polarity agreement are included, so the benchmark leaves out ambiguous sentences and is likely easier than unfiltered Nepali text. Scores on it may not carry over to text in general.
• Narrow set of speakers. All text comes from 11 prominent accounts, in a public and often political register, and the accounts contribute unevenly (94 to 535 sentences each). Results should not be generalised to everyday Nepali social media.
• Class imbalance. Both annotators chose a strong label on few sentences: 63 strongly_positive and 22 strongly_negative.
• Text cleaning artifacts. Script stripping is based on character runs and can split words joined to Latin text. Sentence splitting is simple and does not handle every case, such as quoted speech.
• Uneven time coverage. Tweets are not evenly distributed across the date range.

Citation

If you use this dataset, please cite our preprint:

@misc{baral2026nepsee,
  title         = {{NEPSEE}: A Double-Annotated Benchmark for Nepali Sentiment Analysis},
  author        = {Baral, Gaurab and Nepal, Mimansak and Basnet, Grishma},
  year          = {2026},
  eprint        = {2610.19190},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2610.19190}
}

License

This dataset is released under the MIT License.