samanjoy2/DepressionEmo-SL

Dataset Card for FineGrainedDepressionEmo

Dataset Summary

FineGrainedDepressionEmo is a sentence-level emotion classification dataset derived from long-form, depression-related user posts. Each post is split into sentences and each sentence is annotated with a fine-grained emotion label capturing distinct facets of depressive experience (e.g., sadness, hopelessness, worthlessness, suicide intent).

The dataset is designed for studying fine-grained emotional signals in mental health narratives and for evaluating context-aware models that operate over sentence sequences within a thread.

Supported Tasks and Leaderboards

Primary task: single-label multi-class emotion classification at sentence level.

Given a sentence, the goal is to predict one of the following emotion labels:

• anger
• brain dysfunction
• emptiness
• hopelessness
• loneliness
• sadness
• suicide intent
• worthlessness

There is no official leaderboard yet, but the dataset is suitable for:

• Baseline sentence classification with BERT-like models (C0, no context).
• Context-aware models that include previous/next sentences or full threads (C1–C3 in the accompanying EmoShiftNet code).

Languages

• English (informal, user-generated, often Reddit-style narratives).

Dataset Structure

Data Instances

Each row corresponds to a single sentence from a depression-related narrative.

Example (HF-sanitized version):

item_id,sentence,emotion_final
hhcq6e_1,My mum had a boyfriend when I was around 6 or 7.,sadness
hhcq6e_2,She met him while she was volunteering at a prison.,sadness

Data Fields

In the Hugging Face upload variant, each CSV file has exactly three columns:

• item_id (string): identifier for the sentence. Sentences from the same original post share a common base id, and in the HF-ready copy each sentence is made unique by appending a 1-based index within that thread (e.g., hhcq6e_1, hhcq6e_2, …). The numeric suffix reflects sentence order in the original narrative.
• sentence (string): the sentence text.
• emotion_final (string): the final human-validated emotion label, one of: anger, brain dysfunction, emptiness, hopelessness, loneliness, sadness, suicide intent, worthlessness.

Data Splits

Typical files:

• train_merged_final_emotions_with_final_label.csv
• val_merged_final_emotions_with_final_label.csv
• test_merged_final_emotions_with_final_label.csv
• all_merged_final_emotions_with_final_label.csv

Split sizes depend on the exact configuration, but the combined file (all_merged_final_emotions_with_final_label.csv) contains 32,347 labeled sentences.

Label Distribution

Label counts in the full combined file:

• sadness: 10,023
• hopelessness: 6,494
• loneliness: 5,107
• anger: 4,168
• worthlessness: 2,250
• suicide intent: 1,821
• emptiness: 1,720
• brain dysfunction: 764

The label distribution is moderately imbalanced; macro-averaged metrics (macro F1) are recommended when reporting results.

Text Characteristics

Sentences are mostly well-formed English, often written in first person and describing subjective, depression-related experiences.

End-of-sentence punctuation in the combined file:

• . (period) — 27,507 sentences
• ? (question mark) — 2,260 sentences
• ! (exclamation mark) — 425 sentences

All other final characters (letters, closing brackets, quotes, ellipsis, etc.) are comparatively rare. This is helpful when choosing maximum sequence length and when designing sentence-boundary-aware models.

Data Processing and Annotations

High-level pipeline (see the EmoShiftNet repository for full details):

1. Collect long-form, depression-related posts.
2. Split posts into sentences.
3. Derive a final consensus / manually validated label emotion_final.
4. For the Hugging Face upload:
   • Remove intermediate model-based label columns.
   • Make item_id sentence-unique via a per-thread index suffix.

Usage

In Python with datasets:

from datasets import load_dataset

ds = load_dataset("samanjoy2/FineGrainedDepressionEmo")
df = ds["train"].to_pandas()  # if you define train/val/test splits

Typical modeling steps:

1. Use sentence as input text.
2. Encode emotion_final as class labels.
3. Optionally recover base thread ids by stripping the numeric suffix from item_id (e.g., hhcq6e_1 → hhcq6e) and group sentences to build context-aware inputs (previous/next sentences, full thread, etc.).

Ethical Considerations

The dataset originates from user-generated content describing mental health and depression-related experiences.

Users should:

• Avoid attempts to identify or deanonymize individuals.
• Avoid deploying models trained on this data in high-stakes clinical settings without appropriate oversight.
• Clearly communicate limitations and potential biases of any models trained on this dataset.

Citation

@inproceedings{joy2026context,
  author    = {Joy, Saman Sarker and Aishi, Tanusree Das and Anwar, Shuchismita and Riju, Tanjim Islam and Mahmood, Adnan},
  title     = {Context as a Signal: Context-Aware Transformers for Fine-Grained Depression Emotions},
  booktitle = {2026 IEEE International Conference on Digital Health (ICDH)},
  year      = {2026},
  publisher = {IEEE},
}