README.md

metadata

license: apache-2.0
language:
  - en
task_categories:
  - text-classification
task_ids:
  - multi-label-classification
tags:
  - emotion
  - twitter
  - affect
pretty_name: 'SemEval-2018 Task 1: Affect in Tweets'
size_categories:
  - 1K<n<10K
dataset_info:
  features:
    - name: text
      dtype: string
    - name: labels
      dtype: sequence
      feature:
        dtype: int32
  splits:
    - name: train
    - name: validation
    - name: test

SemEval-2018 Task 1 (Preprocessed)

Dataset Description

This dataset contains a preprocessed and standardized version of SemEval-2018 Task 1: Affect in Tweets for multi-label emotion classification.

The original SemEval task focuses on emotion detection in tweets.
This version has been adapted to support multi-label learning and aligned with a unified emotion encoding scheme used across multiple benchmark datasets in this project.

---

Supported Tasks

• Multi-label emotion classification
• Emotion analysis in short social media texts
• Cross-dataset benchmarking
• Emotion representation learning

---

Dataset Structure

The dataset is split into:

• train
• validation
• test

All splits follow the same schema.

---

Data Format

Each example consists of:

• text (string): Preprocessed tweet text
• labels : Multi-one-hot encoded emotion labels

The label vector is 28-dimensional to maintain compatibility with other datasets.
Only 11 emotions are present in SemEval; all other emotion positions are set to 0.

Each label is binary:

• 1 → emotion present
• 0 → emotion absent

Multiple emotions may be active for a single sample.

---

Emotion Label Mapping

SemEval Emotion Set (11 Emotions)

Index	Emotion
0	Anger
1	Anticipation
2	Disgust
3	Fear
4	Joy
5	Love
6	Optimism
7	Pessimism
8	Sadness
9	Surprise
10	Trust

Unified Encoding Note

To support cross-dataset training and evaluation, SemEval labels are embedded into a 28-class emotion space.
Emotion classes not present in SemEval are encoded as absent (0).

---

Preprocessing Details

The following preprocessing steps were applied:

• Conversion to multi-one-hot label encoding
• Mapping to a unified 11-class emotion space
• Removal of unused metadata and tweet-specific fields
• Text normalization
• Preprocessing applied prior to tokenization

---

Intended Use

This dataset is intended for:

• Training and evaluating multi-label emotion classifiers
• Emotion analysis of social media content
• Cross-dataset generalization experiments
• Benchmarking emotion representations

---

Limitations

• The dataset contains preprocessed text only
• Raw SemEval data is not included
• Tweets may contain noise, slang, or informal language
• Emotion annotations reflect annotator perception and task-specific definitions

---

Citation

If you use this dataset, please cite the original SemEval-2018 Task 1 paper:

@inproceedings{SemEval2018Task1,
  author    = {Mohammad, Saif M. and Bravo-Marquez, Felipe and Salameh, Mohammad and Kiritchenko, Svetlana},
  title     = {SemEval-2018 {T}ask 1: {A}ffect in Tweets},
  booktitle = {Proceedings of the International Workshop on Semantic Evaluation (SemEval-2018)},
  address   = {New Orleans, LA, USA},
  year      = {2018}
}

also cite our paper:
@article{DualObjectivesEmotion2026,
  author  = {Arnab Karmakar, Subinoy Bera}
  title   = {Do We Need a Classifier? Dual Objectives Go Beyond Baselines in Fine-Grained Emotion Classification},
  year    = {2026},
  journal = {Research Gate},
  doi     = {10.13140/RG.2.2.16084.46728},
  url     = {https://www.researchgate.net/publication/399329430_Do_We_Need_a_Classifier_Dual_Objectives_Go_Beyond_Baselines_in_Fine-Grained_Emotion_Classification}
}