This dataset contains transcripts of real mental-health therapy sessions and is shared strictly for non-commercial academic research.
By requesting access you agree to: (1) use the data only for non-commercial research purposes, (2) not attempt to re-identify any individual in the dataset, (3) store the data securely and not redistribute it without permission, (4) cite the original work in any publication that uses this dataset, (5) not use this dataset for any clinical, diagnostic, or therapeutic application.

IndieMH: Code-Mixed Mental Health Counseling Conversations

Dataset Summary

IndieMH is an extended, enriched version of the publicly available HOPE dataset (published at WSDM 2022). It augments the original HOPE counseling transcripts with two new layers of annotation introduced in our LREC 2026 paper: code-mixed Hindi–English renderings of every utterance and multi-label emotion annotations (up to four co-occurring emotions per turn). The structural annotations carried forward from HOPE — speaker role, dialogue acts, and conversation sub-topics — are described in the HOPE dataset.

This dataset was introduced in:

Knowledge-Infused Hierarchy-Aware Emotion Recognition in Code-mixed Mental Health Counseling Conversations Aseem Srivastava, Kushagra Mittal, Anusha Tiwari and Md. Shad Akhtar Proceedings of LREC 2026 Paper link and citation: to be released

Dataset Description

Background

Effective counseling is often best achieved in a client's preferred language, allowing better emotional resonance. Despite this, most existing research in emotion recognition in counseling focuses predominantly on English, overlooking the rich emotional and linguistic complexities of other widely spoken languages. Hinglish — a code-mixed blend of Hindi and English — is one such underexplored linguistic medium that millions use to express their emotions authentically.

To address this gap, IndieMH lays a foundational step in developing a mental-health conversation dataset in code-mixed Hinglish. Counseling conversations from publicly available sources (the HOPE dataset, WSDM 2022) are manually translated into Hinglish. The dataset is employed for the emotion classification task for counseling patients, annotated with 13 emotional states organised under 3 broad emotion categories following an exhaustive annotation guideline. Rigorous sanity checks ensure that the quality of IndieMH adheres to research standards.

Alongside the dataset, we propose Healer — a novel knowledge-cum-hierarchy aware method for counseling emotion classification in the Hinglish language — benchmarked against 11 baseline methods using accuracy, weighted-F1, and weighted-precision.

Relationship to the HOPE Dataset

	HOPE (WSDM 2022)	IndieMH (LREC 2026)
Utterance transcripts	✅	✅ (carried forward)
Speaker labels (T / P)	✅	✅ (carried forward)
Dialogue act labels	✅	✅ (carried forward)
Sub-topic labels	✅	✅ (carried forward)
Code-mixed renderings	❌	✅ new
Emotion labels (multi-label)	❌	✅ new

The HOPE dataset is available via access request at: https://github.com/LCS2-IIITD/SPARTA_WSDM2022#hope-dataset-access-request

Languages

Language	Role
English	Primary utterance language
Hindi (Devanagari & Romanized)	Code-mixed renderings

Dataset Structure

Splits

Split	Conversations	Utterances	Avg. turns/conv
train	130	8,322	64.0
test	39	2,010	51.5
validation	21	1,191	56.7
total	190	11,523	60.6

Data Fields

Field	Type	Source	Description
`ID`	string	HOPE	Unique turn identifier in `{conversation_id}_{turn_index}` format (e.g. `108_0`)
`Speaker`	string	HOPE	`T` = Therapist · `P` = Patient
`Utterance`	string	HOPE	Original English utterance
`Dialogue_Act`	string	HOPE	Dialogue act label (see table below)
`Sub topic`	string	HOPE	Conversation sub-topic label (see table below)
`Code-Mix`	string	IndieMH	Code-mixed Hindi–English rendering of the utterance (new in this work)
`Emotion`	string	IndieMH	Comma-separated emotion label(s); up to four co-occurring labels per utterance (new in this work)

Annotation Schema

Emotion Labels

Utterances may carry up to four co-occurring emotion labels drawn from a Plutchik-inspired taxonomy:

Neutral · Sadness · Scared · Annoyance · Confusion · Anticipation · Contempt · Joy · Trust · Serenity · Disgust · Anger · Surprise

Emotion	Occurrences	% of labelled instances
Neutral	8,960	73.8%
Sadness	1,222	10.1%
Scared	568	4.7%
Annoyance	364	3.0%
Confusion	339	2.8%
Anticipation	202	1.7%
Contempt	98	0.8%
Joy	92	0.8%
Others	294	2.4%

Dialogue Act Labels

Label	Description
`id`	Information delivery
`irq`	Information request
`gc`	General chat
`crq`	Clarification request
`cd`	Clarification delivery
`ack`	Acknowledgement
`yq`	Yes/no question
`gt`	Greeting
`on`	Opinion/narrative
`od`	Opinion delivery
`op`	Opinion/perspective
`orq`	Opinion request
`cr`	Confirmation request
`cv`	Confirmation
`ci`	Closing

Sub-Topic Labels

Label	Description	% of utterances
`routine`	Day-to-day activities and coping strategies	47.1%
`inactive`	Opening/closing exchanges, off-topic turns	21.4%
`symp/reasoning`	Symptom discussion and clinical reasoning	20.6%
`story`	Personal narrative and life events	10.8%

Dataset Statistics

Metric	Value
Total utterances	11,523
Unique conversations	190
Avg. utterance length (words)	22.0
Median utterance length (words)	11
Max utterance length (words)	799
Speaker balance (T / P)	50.4% / 49.6%
Utterances with ≥2 emotion labels	554 (4.8%)

Access

This dataset is gated. The dataset page and sample previews are publicly visible. To download the full data, you must agree to the terms of use. Access is granted automatically upon agreement.

Citation

If you use IndieMH, please cite both this work and the original HOPE dataset it extends.

This work (LREC 2026):

Paper link and BibTeX citation will be released upon publication.

Aseem Srivastava, Kushagra Mittal, Anusha Tiwari and Md. Shad Akhtar. "Knowledge-Infused Hierarchy-Aware Emotion Recognition in Code-mixed Mental Health Counseling Conversations." Proceedings of LREC 2026.

Original HOPE dataset (WSDM 2022):

Please refer to the HOPE dataset repository for the BibTeX entry of the original WSDM 2022 paper.

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

Downloads last month: 11

Total file size:

1.61 MB