README.md

metadata

license: cc-by-4.0
language:
  - ur
  - en
tags:
  - roman-urdu
  - emotion-classification
  - text-classification
  - ekman
  - nlp
  - low-resource-nlp
  - affective-computing
  - social-media
  - urdu-nlp
  - mental-health
  - whatsapp
  - inter-annotator-agreement
pretty_name: RUEmoCorp — Roman Urdu Emotion Corpus
size_categories:
  - 10K<n<100K
task_categories:
  - text-classification
task_ids:
  - multi-class-classification
annotations_creators:
  - expert-generated
  - machine-generated
language_creators:
  - found
multilinguality:
  - monolingual
source_datasets:
  - Khubaib01/RomanUrdu-NLP-Sentiment-Corpus
configs:
  - config_name: ruemocorp-annotated
    data_files:
      - split: train
        path: RUEmoCorpus.csv
  - config_name: ruemocorp-silver
    data_files:
      - split: train
        path: RUEmoCorp_134k_silver/RUEmoCorp_134_labeled.csv

RUEmoCorp

The largest publicly available, human-annotated, inter-annotator-agreement-validated emotion dataset for Roman Urdu.

---

Table of Contents

1. Dataset Overview
2. Background and Motivation
3. Dataset Statistics
4. Inter-Annotator Agreement (IAA)
5. Annotation Methodology
6. Annotation Team
7. Data Fields
8. Data Splits
9. Source and Collection
10. Associated Model
11. Related Resources
12. Datasheet (Gebru et al., 2018)
13. Citation
14. License and Ethics
15. Contact

---

Dataset Overview

RUEmoCorp (Roman Urdu Emotion Corpus) is a large-scale, manually curated, expert-annotated dataset of Roman Urdu social media and conversational texts labeled across 7 emotion categories: joy, anger, sadness, fear, disgust, surprise, and none. It is the training corpus behind roman-urdu-emotion-xlmr-v2 — the highest-accuracy open-source emotion classifier for Roman Urdu, achieving Macro F1 = 0.9896.

Data was collected from Pakistani social media platforms and WhatsApp conversations and underwent a rigorous multi-phase annotation process by four expert annotators recruited from three independent Pakistani universities. An inter-annotator agreement (IAA) study on a 700-sample benchmark yields Fleiss' κ = 0.6588 and Mean Pairwise Cohen's κ = 0.6597, indicating substantial agreement (Landis & Koch, 1977) — a strong result for a 7-class affective labeling task in a low-resource, orthographically irregular language.

The emotion taxonomy adopts Ekman's six universal basic emotions augmented with a none class for emotionally neutral utterances — a deliberate design choice absent from prior Roman Urdu emotion work, which has used only four or six categories. Omitting a neutral class forces classifiers to assign emotional labels to neutral text, inflating false positive rates in deployed systems.

This dataset fills a documented gap: prior to this release, no large-scale, openly accessible, IAA-validated emotion corpus existed for Roman Urdu, despite Roman Urdu being the dominant digital writing mode for over 230 million Urdu speakers worldwide. RUEmoCorp is permanently archived on Harvard Dataverse (doi:10.7910/DVN/BPWHOZ) and released under CC BY 4.0.

2. Background and Motivation

2.1 The Roman Urdu Digital Language Problem

Urdu is the national language of Pakistan and a major language of India, with over 230 million speakers. However, in digital communication — social media, messaging apps, online forums — native speakers overwhelmingly write in Roman Urdu: Urdu lexicon and grammar rendered in the Latin script, without standardised orthography.

This creates a profound NLP challenge:

• The same word can be spelled in dozens of valid ways (khushi, khushee, khushi, khuushi)
• No standard keyboard layout, no spell-checker, no official romanisation standard
• Extensive code-switching with English at both the word and phrase level
• Existing Urdu NLP resources built for Nastaliq script do not transfer to Roman Urdu

2.2 Why Emotion Classification

Emotion classification is foundational for downstream applications in mental health monitoring, social media analysis, customer feedback systems, and conflict detection in multilingual communities. For Roman Urdu specifically, no validated emotion resource existed before this work.

2.3 Research Lineage

This dataset is part of a growing research programme on Roman Urdu affective computing:

Resource	Size	Task	Status
RomanUrdu-NLP-Sentiment-Corpus	134K	3-class sentiment	Public
roman-urdu-sentiment-xlm-r	—	Sentiment model	Public
RUEC-28K (this dataset)	28K	7-class emotion	Public
roman-urdu-emotion-xlmr-v2	—	Emotion model	Public
RomanUrdu-NLP-Emotion-Corpus-134K	134K	Emotion (model-labeled)	Forthcoming

---

3. Dataset Statistics

RUEmoCorp (28k) training

3.1 Size and Format

Property	Value
Total samples	28,000
Emotion classes	7
Annotation format	Single label per sample
Language	Roman Urdu (code-switched with English)
Script	Latin (Roman)
Domain	Social media text
Format	CSV / Parquet

3.2 Class Distribution

Emotion Label	Sample Count	% of Dataset
Happy	~4,000	~14.3%
Sad	~4,000	~14.3%
Anger	~4,000	~14.3%
Disgust	~4,000	~14.3%
Fear	~4,000	~14.3%
Surprise	~4,000	~14.3%
Neutral	~4,000	~14.3%

The dataset was constructed with approximate class balance to ensure unbiased classifier training.

Dataset Statistics — RUEmoCorp-silver

RUEmoCorp is annotated by the Khubaib01/roman-urdu-emotion-xlmr-v2

Overview

Property	Value
Total utterances	134,053
Annotation method	Automated — roman-urdu-emotion-xlmr-v2
Confidence threshold	≥ 0.75 (softmax probability)
Mean confidence	0.8039
Median confidence	0.8733
Mean prediction entropy	0.7623
Low-confidence rows (< 0.75)	10,109 (7.54%)
Fallback / unresolved rows	0 (0.00%)

The high median confidence (0.8733) indicates that the majority of retained predictions are well above the retention threshold, with low-confidence rows constituting only 7.54% of the corpus. Zero fallback rows confirm complete model coverage across all retained utterances.

---

Class Distribution (with 95% Wilson Confidence Intervals)

Emotion	Count	%	CI Lower	CI Upper
joy	28,389	21.18%	0.2096	0.2140
none	28,167	21.01%	0.2079	0.2123
disgust	25,959	19.36%	0.1915	0.1958
sadness	22,570	16.84%	0.1664	0.1704
anger	18,275	13.63%	0.1345	0.1382
fear	6,613	4.93%	0.0482	0.0505
surprise	4,080	3.04%	0.0295	0.0314

⚠️ The distribution is naturally imbalanced, reflecting the organic frequency of emotional expression in scraped social media and WhatsApp data. joy and none together account for ~42% of the corpus. fear and surprise are the least frequent classes (combined ~8%). Users should apply class reweighting or stratified sampling before using this corpus as a primary training source.

---

Per-Class Confidence Statistics

Emotion	Mean Conf.	Std	Median Conf.	Min	Max
anger	0.8175	0.1292	0.8778	0.2291	0.9105
disgust	0.7819	0.1499	0.8630	0.1875	0.9142
fear	0.7545	0.1842	0.8484	0.1900	0.9320
joy	0.8476	0.1244	0.9037	0.2336	0.9290
none	0.8155	0.1361	0.8804	0.2207	0.9193
sadness	0.7696	0.1510	0.8488	0.2147	0.9068
surprise	0.7699	0.1755	0.8637	0.2195	0.9294

joy and anger record the highest mean confidence (0.8476 and 0.8175 respectively), consistent with their strong per-class F1 scores on the human-annotated gold set. fear and surprise record the lowest mean confidence and highest standard deviation, reflecting their lower corpus frequency and greater lexical ambiguity in informal Roman Urdu — also the classes with the widest Wilson CI bounds in the distribution table above. All per-class median confidence values exceed 0.84, indicating that the central tendency of predictions is substantially above the 0.75 retention threshold across all seven categories.

3.3 IAA Validation Subset

A stratified random sample of 700 instances (100 per class) was independently re-annotated by all four annotators for the IAA study. Results are reported in Section 4.

---

4. Inter-Annotator Agreement (IAA)

IAA was computed on a 700-sample stratified subset independently re-annotated by all four members of the annotation team. All annotators worked blindly — they had no access to the original labels or each other's responses during annotation.

4.1 Aggregate Agreement

Metric	Value	Interpretation
Fleiss' Kappa (κ)	0.6588	Substantial agreement
Mean Pairwise Cohen's κ	0.6597	Substantial agreement
Total IAA Samples	700	Stratified (100 per class)
Full Agreement (4/4 annotators)	348 (49.7%)	—
Majority Agreement (3/4 annotators)	241 (34.4%)	—
Ambiguous (no majority)	111 (15.9%)	—

Benchmark context: A Fleiss' κ of 0.659 is considered substantial agreement on the Landis & Koch (1977) scale (0.61–0.80 = substantial). For a 7-class affective labeling task in a low-resource, orthographically irregular language, this result compares favourably with comparable published corpora. SemEval-2018 Task 1 reported average κ values in the 0.60–0.72 range for multi-label emotion classification in English tweets.

4.2 Agreement Breakdown by Sample Category

Agreement Category	Count	Percentage
Full agreement (all 4 annotators agree)	348	49.7%
Majority agreement (3 of 4 agree)	241	34.4%
Ambiguous (2-2 split or no clear majority)	111	15.9%
Total	700	100%

4.3 IAA Visualisation

Figure 1 — IAA Agreement Distribution Chart

4.4 Ambiguous Sample Handling

The 111 ambiguous samples (15.9%) — those where no majority label emerged — were adjudicated by the corresponding author (Khubaib Ahmad) using the following protocol:

1. Review the sample in its original posting context
2. Apply the primary criterion: which emotion would a native Roman Urdu speaker most likely intend?
3. If genuinely unresolvable, the sample was marked with the label most consistent with the broader textual context
4. Edge cases between Anger and Disgust are the dominant source of ambiguity — these two classes share lexical overlap in Roman Urdu informal expression and represent the known hard boundary in affective computing for South Asian languages

---

5. Annotation Methodology

5.1 Annotation Design

The annotation followed a three-phase blind re-annotation protocol designed to maximise label reliability for difficult low-resource boundary cases:

Phase 1 — Independent Annotation Each annotator labeled all assigned samples independently, with no communication between annotators. Annotation guidelines were provided in written form and discussed in a single calibration session before annotation began.

Phase 2 — Calibration on Boundary Cases After Phase 1, all annotators jointly reviewed a set of 35 pre-selected boundary-case samples (primarily anger/disgust and sad/neutral pairs). The purpose was alignment on class definitions, not correction of existing labels. Phase 2 outputs were not used in the final dataset.

Phase 3 — Re-annotation of High-Disagreement Samples Samples flagged as high-disagreement from Phase 1 were re-annotated independently by all four annotators. Final labels were determined by majority vote.

5.2 Emotion Label Definitions

Annotators were provided with the following operational definitions, grounded in Ekman's (1992) six basic emotions framework, adapted for the Roman Urdu social media context:

Label	Definition	Roman Urdu Signal Examples
Happy	Joy, contentment, excitement, celebration	maza aa gaya, khushi ho rahi, zabardast
Sad	Grief, disappointment, longing, loss	dil dukha, rona aa raha, yaad aa rahi
Anger	Frustration, rage, strong displeasure directed outward	gussa aa raha, bura lag raha, tang aa gaya
Disgust	Revulsion, moral rejection, strong aversion	nafrat hai, ganda lagta, sharm karo
Fear	Anxiety, dread, nervousness about outcome	dar lag raha, fikr ho rahi, kuch bura hoga
Surprise	Unexpected reaction, shock (positive or negative)	hairan reh gaya, pata nahi tha, achanak
Neutral	No dominant emotion detectable; informational or factual	news sharing, plain description, announcements

5.3 Annotation Challenges Specific to Roman Urdu

Several properties of Roman Urdu text created annotation challenges not present in standard NLP annotation tasks:

• Orthographic variability: The same word in different spellings was sometimes perceived differently by annotators. Guidelines included canonical forms.
• Code-switching: English emotion words embedded in Roman Urdu phrases (e.g., "itna sad feel ho raha") required consistent treatment. Guidelines specified to treat code-switched expressions at their semantic value.
• Implicit emotion: Roman Urdu social media text frequently expresses emotion indirectly through cultural references, humour, or rhetorical questions. These samples constituted the majority of ambiguous cases.
• Anger-Disgust boundary: The most frequent source of disagreement. Both emotions share vocabulary in informal Pakistani social media usage. The calibration session (Phase 2) focused specifically on this boundary.

---

6. Annotation Team

RUEC-28K was annotated by a dedicated four-person expert team, all native or fluent Roman Urdu speakers with academic backgrounds in relevant fields.

Annotator	Affiliation	Location	Role
Muzammil Shadab	Bahauddin Zakariya University (BZU)	Multan, Punjab	Annotator
Sara	COMSATS University Islamabad (CUI)	Islamabad	Annotator
Faiez Ahmad	Emerson University Multan (EUM)	Multan, Punjab	Annotator
Khadija Faisal	Emerson University Multan (EUM)	Multan, Punjab	Data Manager & Annotator

Corresponding author / Project lead: Muhammad Khubaib Ahmad, Emerson University Multan (EUM), Multan, Punjab.

All annotators participated in the calibration session and the IAA study. The annotation team has no financial conflict of interest in the publication of this dataset.

---

7. Data Fields

{
  "message":       str,   # Raw Roman Urdu text (social media post or message)
  "emotion_label": str    # One of: anger | disgust | fear | happy | neutral | sad | surprise
}

Field Details

message

• Raw Roman Urdu text, preserved as collected with minimal preprocessing
• May contain code-switched English words or phrases
• May contain common social media abbreviations and informal orthography
• No personally identifiable information (PII) — all samples were anonymised prior to release
• Length: typically 5–80 tokens

emotion_label

• String label, lowercase
• Assigned by majority vote across 4 expert annotators for the IAA-validated subset
• For the full 28K corpus: assigned by primary annotator and reviewed by data manager (Khadija Faisal)
• Valid values: anger, disgust, fear, happy, neutral, sad, surprise

---

8. Data Splits

Split	Size	Notes
Train	~24,000	Used to train roman-urdu-emotion-xlmr-v1 and v2
Validation	~2,000	Held out during training
Test	~2,000	Used for final evaluation; annotated by same team

Note on test set construction: The test set was sampled from the same 28K corpus and labeled by the same annotation team as the training set. This is a known limitation of the current release — see Section 12.4 (Limitations). An independently annotated external validation set is in preparation.

---

9. Source and Collection

9.1 Parent Corpus

RUEC-28K is a subset of the RomanUrdu-NLP-Sentiment-Corpus (134K samples), which was collected from publicly accessible Pakistani social media platforms. Texts were selected to represent diverse emotional expression in everyday Roman Urdu communication.

9.2 Selection Criteria for RUEC-28K

From the 134K sentiment corpus, 28K samples were selected for emotion annotation based on:

• Sufficient emotional signal (low-emotion purely informational text excluded)
• Reasonable length (very short texts < 4 tokens and very long texts > 200 tokens excluded)
• Linguistic diversity (orthographic variants represented across classes)
• Approximate class balance across 7 emotion categories

9.3 Preprocessing

• User mentions and names replaced with PERSON
• URLs removed
• Duplicate samples removed
• No stemming, lemmatisation, or normalisation applied — raw orthographic variety preserved intentionally

---

10. Associated Model

RUEC-28K is the training corpus for the roman-urdu-emotion-xlmr-v2 model.

• 🤗 Model: Khubaib01/roman-urdu-emotion-xlmr-v2
• Architecture: XLM-RoBERTa base + two-layer MLP classification head
• Training lineage: Sentiment fine-tuned → Emotion v1 → Emotion v2
• Test set Macro F1: 0.9896
• Paper: in-progress

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="Khubaib01/roman-urdu-emotion-xlmr-v2"
)

result = classifier("yaar dil bht dukha aaj")
print(result)
# [{'label': 'sad', 'score': 0.987}]

---

11. Related Resources

Resource	Description	Link
RomanUrdu-NLP-Sentiment-Corpus	134K sentiment-labeled Roman Urdu corpus	HuggingFace
roman-urdu-sentiment-xlm-r	Sentiment classifier (3-class)	HuggingFace
roman-urdu-emotion-xlmr-v1	Emotion classifier v1	HuggingFace
roman-urdu-emotion-xlmr-v2	Emotion classifier v2 (current best)	HuggingFace
Paper	In-progess	in-progress
Harvard Dataverse	Archival deposit	under-review
RomanUrdu-NLP-Emotion-Corpus-134K	134K model-labeled emotion corpus	HuggingFace

---

12. Datasheet (Gebru et al., 2018)

This datasheet follows the framework proposed by Gebru et al. (2018) — Datasheets for Datasets.

12.1 Motivation

For what purpose was the dataset created? To address the complete absence of large-scale, human-annotated, inter-annotator-agreement-validated emotion data for Roman Urdu — the dominant digital writing mode for Urdu speakers across Pakistan and India.

Who created the dataset and on whose behalf? Muhammad Khubaib Ahmad (AI Research Engineer, Emerson University Multan), as an independent research project, with annotation support from a four-person expert team (see Section 6). No external funding or institutional commission.

Any other comments? This dataset is part of a broader research programme on Roman Urdu affective computing. The associated 134K sentiment corpus and emotion classification models are co-released.

---

12.2 Composition

What do the instances represent? Each instance is a single Roman Urdu social media text — a post, comment, or message — labeled with a single dominant emotion.

How many instances? 28,000 total. Approximately 4,000 per class across 7 emotion categories.

Does the dataset contain all possible instances or a sample? A sample. The parent corpus (134K) was itself a sample of available Roman Urdu social media text.

Is there a label or target associated with each instance? Yes. Each instance has one emotion_label from: anger, disgust, fear, happy, neutral, sad, surprise.

Is any information missing from individual instances? Metadata such as posting timestamp, platform, user demographics are not included to protect privacy.

Are relationships between instances made explicit? No explicit relationships. Instances are treated as independent.

Are there recommended data splits? Yes — see Section 8.

Are there any errors, sources of noise, or redundancies? The 111 ambiguous samples in the IAA study (15.9% of the 700-sample subset) represent genuine annotation uncertainty, primarily at the anger-disgust boundary. These samples are included in the corpus with adjudicated labels. Near-duplicate samples were removed during preprocessing.

Is the dataset self-contained? Yes. All required data is in this repository. The associated models are separately hosted on HuggingFace.

Does the dataset contain data that might be considered confidential? No. All samples were collected from publicly accessible platforms. PII was removed.

---

12.3 Collection Process

How was data associated with each instance acquired? Text collected from publicly accessible Pakistani social media platforms. Emotion labels assigned by expert human annotators following a structured multi-phase protocol (see Section 5).

What mechanisms were used to collect data? Manual collection and curation from public sources. No automated scraping APIs are disclosed in this release.

Over what timeframe was data collected? Collected and annotated over approximately 12 months.

Were any ethical review processes conducted? The project was conducted as independent academic research. All data was sourced from publicly accessible platforms. No PII was retained.

---

12.4 Preprocessing, Cleaning, Labeling

Was any preprocessing/cleaning done? Yes — see Section 9.3. User mentions replaced, URLs removed, duplicates removed. Raw orthographic variety was preserved intentionally.

Was the raw data saved in addition to the preprocessed data? The preprocessed form is the release form. Original raw collection is retained by the corresponding author.

Is the labeling/annotation described in detail? Yes — see Sections 4 and 5.

Was any human labeling conducted? Yes. Four expert annotators (see Section 6). IAA computed on 700-sample stratified subset (Fleiss' κ = 0.659).

---

12.5 Uses

Has the dataset been used for any tasks already? Yes. It is the training corpus for roman-urdu-emotion-xlmr-v2, achieving Macro F1 = 0.9896 on the in-distribution test set.

What are the recommended uses?

• Training and benchmarking Roman Urdu emotion classifiers
• Low-resource multilingual NLP research
• Transfer learning experiments for South Asian languages
• Affective computing and sentiment analysis research

What are the uses that should be avoided?

• Clinical or medical applications without additional validation
• Real-time surveillance or monitoring of individuals
• Applications targeting vulnerable populations without appropriate ethical review
• Any use that relies on the model's output as a ground truth for individual emotional state

---

12.6 Distribution

How is the dataset distributed? Via HuggingFace Datasets (primary) and Harvard Dataverse (archival).

Is the dataset distributed under a copyright or license? Apache 2.0 License. Free for academic and commercial use with attribution.

Have any third parties imposed IP-based restrictions? No.

---

12.7 Maintenance

Who maintains the dataset? Muhammad Khubaib Ahmad. Contact via HuggingFace or the email in Section 15.

Will the dataset be updated? An extended 134K model-labeled version is planned for release. The 28K manually annotated corpus is considered stable.

Will older versions be maintained? Yes. Versioned releases on both HuggingFace and Harvard Dataverse.

---

12.8 Limitations

1. Test set annotator overlap: The test split was annotated by the same team as the training split. In-distribution performance (Macro F1 = 0.9896) should be interpreted accordingly. An externally annotated validation set is in preparation.

2. Domain specificity: Samples are drawn from social media text. Performance on formal text, news, or other domains may differ.

3. Orthographic coverage: While orthographic variety is preserved, the corpus cannot cover all possible romanisation patterns for all Urdu words.

4. Geographic bias: Pakistani Roman Urdu predominates. Indian Roman Urdu may show stylistic and lexical differences.

5. Sarcasm and irony: Implicitly expressed emotions, particularly sarcastic positivity, are a known weak point. These cases appear disproportionately in the ambiguous sample pool.

6. Static snapshot: Social media language evolves. Newer slang or expression patterns post-collection may not be represented.

---

13. Citation

If you use this dataset in your research, please cite:

@data{DVN/BPWHOZ_2026,
author = {Ahmad, Muhammad Khubaib Ahmad and Khadija Faisal},
publisher = {Harvard Dataverse},
title = {{RUEmoCorp}},
UNF = {UNF:6:h03jo4SJGEAKuZCik1R/Bw==},
year = {2026},
version = {V1},
doi = {10.7910/DVN/BPWHOZ},
url = {https://doi.org/10.7910/DVN/BPWHOZ}
}

If you use the associated model, please also cite:

@misc{muhammad_khubaib_ahmad_2026,
    author       = { Muhammad Khubaib Ahmad and Khadija Faisal },
    title        = { roman-urdu-emotion-xlmr-v2 (Revision 7cd7dd2) },
    year         = 2026,
    url          = { https://huggingface.co/Khubaib01/roman-urdu-emotion-xlmr-v2 },
    doi          = { 10.57967/hf/8347 },
    publisher    = { Hugging Face }
}

---

14. Team and Contributions

Name	Role	Affiliation
Muhammad Khubaib Ahmad	Core Researcher · Lead Engineer · Project Administration · Model Development	Independent Researcher
Khadija Faisal	Data Manager · Annotation Coordination · Annotator	Emerson University Multan
Muzammil Shadab	Annotator	Bahauddin Zakariya University, Multan
Sara	Annotator	COMSATS University Islamabad
Faiez Ahmad	Annotator	Emerson University Multan

---

15. License and Ethics

License: Apache 2.0

This dataset is freely available for academic research, commercial use, and derivative works with appropriate attribution.

Ethical considerations:

• All source texts were collected from publicly accessible platforms
• No personally identifiable information (PII) is present in the released dataset
• The emotion labels reflect the expressed emotion in text as interpreted by expert annotators — they do not constitute claims about the psychological state of any individual
• Emotion classification systems carry inherent risks of misuse, particularly in surveillance, profiling, or targeting applications. Users of this dataset are responsible for ensuring their applications comply with applicable data protection laws and ethical guidelines
• The annotation team was compensated appropriately for their work

---

16. Contact

Muhammad Khubaib Ahmad AI Research Engineer Multan, Punjab, Pakistan

• 🤗 HuggingFace: Khubaib01
• 📄 Paper: In-progress

For questions about the dataset, annotation methodology, or collaboration requests, please open a Discussion on this repository.

---

RUEmoCorp — The largest publicly available, IAA-validated Roman Urdu emotion dataset. Released to support low-resource NLP research for South Asian languages.