FaisalHejary/go-emotions-multilingual-embedding

Multilingual GoEmotions for embedding training

Two classification tasks for training and evaluating instruction-following embedding models, built from two machine translations of GoEmotions (Reddit comments labelled with 28 emotions) in 36 languages.

Every row is a query with one or more correct labels (positives) and wrong ones (negatives), all written as text, so a model learns to pick the label most similar to the query.

Tasks

Task	Query	Positives	Negatives
emotion	A comment	Its emotions, one or more of 28 ("neutral" included)	The 4 wrong emotions most similar to the comment
sentiment	A comment, not in Arabic or English	"positive", "neutral" or "negative"	The other 2

Task	Query instruction	Document instruction
emotion	Identify the emotions expressed in this text, or neutral if it expresses none	Represent this emotion for classification
sentiment	Identify whether this text expresses a positive, negative or neutral feeling, or neutral if it expresses none	Represent this sentiment for classification

Emotion labels are the GoEmotions names, in English, whatever the comment's language: 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.

Splits

The original GoEmotions train, validation and test splits.

Task	train	validation	test
emotion	68,214	16,193	16,162
sentiment	33,956	6,854	6,849
Total	102,170	23,047	23,011

• Horizon-Labs translates only GoEmotions train, so validation and test cover the 6 AnasAlokla languages: en, ar, fr, es, tr, de (sentiment: fr, es, tr, de).
• sentiment in train is 15,991 positive, 14,854 negative and 3,111 neutral.
• A comment can be in both tasks.

Columns

Column	Description
task	emotion or sentiment
query_inst	Instruction for embedding the query
query	The comment
document_inst	Instruction for embedding the labels
positives	Correct labels
negatives	Wrong labels

How it was built

1. Load both translations. AnasAlokla/multilingual_go_emotions (English and 5 translations) and Horizon-Labs/go-emotions-multilingual (35 languages, GoEmotions train only). Languages are mapped to ISO codes and labels to GoEmotions names.

2. Back to the original splits. Every row is traced to its GoEmotions comment (the AnasAlokla id minus its 2-letter language prefix; Horizon-Labs' source_index) and placed in that comment's GoEmotions split. AnasAlokla re-split GoEmotions at random (60% of its test comes from GoEmotions train), so its own splits would put translations of test comments in train.

3. Dedup.

   • One translation per comment and language; AnasAlokla wins when both have one.
   • Within a split, texts equal after Unicode normalization, lowercasing and stripping punctuation are kept once, or dropped when their emotions disagree.
   • Train texts also in validation or test are dropped. Empty texts are dropped.

   This gives 605,879 train, 32,283 validation and 32,345 test comments.

4. Rank and margin filter, with Qwen3-VL-Embedding-2B and the emotion instructions. Every emotion is ranked by cosine similarity to the comment.

   • Rows with a correct emotion outside the top 5 are dropped as likely mislabelled.
   • The 4 most similar wrong emotions are kept as negatives.
   • Rows where the weakest correct emotion beats the hardest negative by 0.10 or more are easy: a random 25% of them are kept.

5. emotion: cap 3,000 rows per emotion, drawn at random. A comment with several emotions can pass through a rarer one, so frequent emotions reach about 4,700.

6. sentiment is drawn from the same filtered rows as emotion.

   • Arabic and English comments are dropped.
   • Emotions map to sentiment using the GoEmotions paper's grouping. Its ambiguous emotions (surprise, curiosity, confusion, realization) are left out, and comments whose other emotions disagree in sentiment are dropped.
     • positive: admiration, amusement, approval, caring, desire, excitement, gratitude, joy, love, optimism, pride, relief
     • negative: anger, annoyance, disappointment, disapproval, disgust, embarrassment, fear, grief, nervousness, remorse, sadness
     • neutral: neutral
   • At most 2,000 rows per emotion are kept.
   • Then a sentiment rank and margin filter, as in step 4: the right sentiment must be in the top 2 of 3, and rows with a sentiment margin of 0.10 or more keep a random 25%.

All three splits are filtered, so scores on validation and test are lower than on unfiltered data.

Choosing the instructions

Picked by zero-shot macro F1 with Qwen3-VL-Embedding-2B on test, where the prediction is the label with the highest cosine similarity.

• emotion, on all 32,345 test comments before filtering: 6 query instructions scored macro F1 0.182–0.212. Naming "neutral" won (accuracy 0.381 against at most 0.251). The document instruction raised macro F1 from 0.212 to 0.267 (accuracy 0.388, micro F1 0.358). Writing a description for each emotion ("joy: happiness, delight or pleasure") scored lower than the bare names (macro F1 0.250 at best).
• sentiment, on 9,329 test rows before its own margin filter: no instructions scored macro F1 0.413, the document instruction alone 0.626, and the chosen pair 0.694 (accuracy 0.719).

Limitations

• All non-English text is machine translated, and labels come from the English comment, so a translation can lose or shift the emotion.
• Validation and test cover only 6 languages; the other 30 are in train only.
• Labels and negatives are chosen with one model and filtered by that same model's ranks, so they reflect what that model finds hard. The hardest negatives are where false negatives hide.
• sentiment neutral is under-represented in train (9%), since it maps from one emotion.

Source, licence and attribution

Built from GoEmotions (Apache 2.0) through its translations by Horizon-Labs (Apache 2.0) and AnasAlokla (no licence stated). Texts were filtered, deduplicated, sampled and reformatted into the tasks above.

@inproceedings{demszky2020goemotions,
  title={GoEmotions: A Dataset of Fine-Grained Emotions},
  author={Demszky, Dorottya and Movshovitz-Attias, Dana and Ko, Jeongwoo and Cowen, Alan and Nemade, Gaurav and Ravi, Sujith},
  booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
  year={2020}
}