anggars/mbti-emotion

Multilingual Math Rock Lyrics for Emotion and MBTI Classification

Dataset Summary

This dataset contains 120,060 song-lyric passages annotated with two independent classification targets: a fine-grained emotion label (28 categories based on the GoEmotions taxonomy) and a Myers-Briggs Type Indicator (MBTI) personality type (16 categories). All passages are written in the stylistic register of the math rock and midwest emo genres and span eleven languages.

The corpus was built to train two multilingual text classifiers, anggars/xlm-emotion and anggars/xlm-mbti, which form the textual branch of a broader multimodal emotion-and-personality analysis system.

Composition notice. The overwhelming majority of the passages are synthetic, generated with the Gemma large language model, while a small set of authentic lyrics was retrieved from public sources. See Data Sources for exact counts. The dataset is intended for research and educational use only.

Supported Tasks

• Emotion classification (text-classification): predict one of 28 emotion categories from a lyric passage.
• MBTI personality classification (text-classification): predict one of 16 MBTI personality types from a lyric passage.

The two targets are provided as separate columns and are meant to be trained as two independent single-label classifiers.

Languages

The Text field is multilingual. The approximate distribution is:

Language	Code	Rows	Share
English	en	41,759	34.79%
Indonesian	id	25,176	20.98%
Japanese	ja	13,647	11.37%
Korean	ko	9,974	8.31%
Spanish	es	8,514	7.09%
French	fr	7,019	5.85%
German	de	4,112	3.43%
Chinese	zh	4,080	3.40%
Portuguese	pt	2,890	2.41%
Russian	ru	2,843	2.37%
Italian	it	1	<0.01%
Total		120,060	100%

Dataset Structure

Data Fields

Field	Type	Description
Text	string	The lyric passage.
Emotion	string	One of 28 GoEmotions categories (see below).
MBTI	string	One of 16 MBTI personality types (see below).
Language	string	ISO 639-1 language code of Text.
source	string	Provenance of the row: gemma (synthetic) or genius (organic).
Title	string	Song title (populated for authentic rows; tagged as [Synthetic Data] for generated rows).
Artist	string	Performing artist (populated for authentic rows; tagged as [Synthetic Gemma] for generated rows).

Data Splits

The dataset ships as a single train split. Users are expected to create their own validation and test partitions (for example, a stratified 80/20 split with a fixed seed).

Split	Rows
train	120,060

Class Labels

Emotion (28 categories, GoEmotions taxonomy)

admiration, amusement, anger, annoyance, approval, caring, confusion, curiosity, desire, disappointment, disapproval, disgust, embarrassment, excitement, fear, gratitude, grief, joy, love, nervousness, optimism, pride, realization, relief, remorse, sadness, surprise, neutral

The emotion labels are approximately balanced, with roughly 4,287 to 4,288 rows per category.

MBTI (16 personality types)

INTJ, INTP, ENTJ, ENTP, INFJ, INFP, ENFJ, ENFP, ISTJ, ISFJ, ESTJ, ESFJ, ISTP, ISFP, ESTP, ESFP

The MBTI labels are approximately balanced, with roughly 7,503 to 7,504 rows per type.

Data Sources

Source	Rows	Share	Description
gemma	119,967	99.92%	Passages generated with the Gemma large language model.
genius	93	0.08%	Authentic lyrics retrieved from public sources.
Total	120,060	100%

Data Collection and Processing

1. Authentic lyric retrieval. A limited pool of genuine math rock and midwest emo lyrics was collected from public lyric providers, then cleaned and language-tagged.
2. Synthetic generation. To reach a balanced, large-scale corpus, the remaining passages were generated with the Gemma large language model, conditioned on a target emotion category, a target MBTI type, a target language, and the math rock / midwest emo stylistic register.
3. Data Cleaning & Re-ordering. Unused features (e.g., Style) were dropped. Metadata was normalized, ensuring synthetic data titles and artists are explicitly labeled as [Synthetic Data] and [Synthetic Gemma]. Finally, the dataset was reordered using a custom hierarchy to prioritize authentic lyrics from key math rock / midwest emo bands at the top of the dataset, while synthetic data was pushed to the bottom.

Intended Uses

This dataset is designed for:

• Fine-tuning multilingual transformer classifiers for lyric-level emotion recognition.
• Fine-tuning multilingual transformer classifiers for exploratory MBTI inference from text.
• Research on affective and stylistic modeling of the math rock and midwest emo genres.

Limitations and Biases

• Predominantly synthetic. About 99.9% of the passages are machine-generated rather than authored by human songwriters. Models trained on this corpus may learn artifacts of the generating model instead of genuine lyric phenomena.
• Weak MBTI supervision. MBTI is a self-report typology with limited scientific validity, and the labels here are assigned as exploratory proxies rather than clinically validated annotations. Results should be interpreted cautiously.
• Genre and length skew. Passages are constrained to a narrow stylistic register and may not generalize to other genres or to full-length songs.
• Uneven language coverage. English and Indonesian dominate the corpus, and some languages (for example, Italian) are represented by only a handful of rows.
• Copyright. Authentic passages remain the property of their respective rights holders and are included in limited quantity for research purposes only.

Citation

If you use this dataset, please cite the associated undergraduate thesis project on multimodal emotion and MBTI analysis of math rock music.

License

Released under CC BY-NC 4.0 for non-commercial research and educational use.