FaisalHejary/dz-emotion-embedding

Dz-Emotion (relabelled) for embedding training

Algerian Arabic (Darija) social-media comments labelled with one of Ekman's six emotions, in a label-as-document format, for training text embedding models. Each emotion is used as a document, so the model learns to rank a comment's emotion above the other five.

Built from Houdna-khilouf/Dz-Emotion (paper), 6,000 YouTube, Facebook and Instagram comments. The comments come from that dataset. The emotion labels do not: they were reassigned by Claude, see Relabelling. Use the source dataset if you need its human annotations.

Columns

Column	Description
task	Task name: always emotion.
query_inst	Instruction for the comment: "Given a social media comment in Algerian Arabic dialect, classify the emotion of the writer: anger, disgust, fear, happiness, sadness or surprise".
query	The comment.
document_inst	Instruction for the emotion: "Represent this emotion for classification".
positives	The comment's emotion, as a list of one.
negatives	The 4 hardest of the other five emotions, hardest first.

Emotions are written as anger, disgust, fear, happiness, sadness and surprise.

How it was built

1. Duplicates. The source repeats comments up to 11 times to reach 1,000 per emotion; only 2,801 of its 6,000 rows are distinct. Comments are compared lowercased with punctuation and emoji removed, and duplicates are dropped, as are 6 emoji-only comments and 1 comment labelled with two emotions, leaving 2,794.
2. Split. The source ships one file, so a random 10% (seed 0) became the test split. Duplicates are removed first, so no comment is in both splits.
3. First filter, on the source labels. Every emotion is ranked by cosine similarity to the comment with Qwen/Qwen3-VL-Embedding-2B. Comments whose emotion ranked last were dropped, leaving 2,621.
4. Relabelling. Those 2,621 comments were relabelled (below).
5. Second filter, on the new labels. The same ranking is repeated:
   • A comment is dropped when its emotion ranks last.
   • The 4 most similar other emotions are kept as negatives.
   • Comments the model already gets right by a margin of 0.10 or more teach little, so only a random 25% of them is kept.

Rows per emotion:

Emotion	train	test
anger	438	45
disgust	420	55
fear	284	32
happiness	440	52
sadness	453	51
surprise	243	29
Total	2,278	264

Unlike the source, the emotions are not balanced: many comments the source calls fear read as grief, and many it calls surprise read as indignation.

Relabelling

Each of the 2,621 comments was labelled by a Claude subagent, about 100 comments per subagent, without seeing the source label. The subagents worked from written definitions of the six emotions and notes on Darija words whose meaning differs from standard Arabic, for example غاظني ("I feel sorry for", not "it angered me"). Each picked the writer's dominant emotion and could name a second one that fits equally well.

Agreement with the source labels:

Measure	Value
Same emotion	72.1%
Cohen's kappa	0.664
Same emotion, or the source's emotion named as the second choice	85.1%

Agreement by source emotion: happiness 96.8%, sadness 82.5%, fear 68.2%, disgust 64.9%, anger 60.8%, surprise 59.4%. The largest disagreements:

Source label	Relabelled as	Comments
anger	disgust (contempt, mockery)	114
surprise	anger (indignant "why ???" questions)	63
disgust	anger	62
fear	sadness (mourning)	55
disgust	happiness (playful teasing)	52

A separate hand check of 200 random comments found 6% of source labels clearly wrong and 18% debatable. The subagents disagreed with the source on 11 of the 12 clearly wrong ones and agreed on 149 of the 152 judged correct.

relabel/dz_emotion_relabel.csv lists every relabelled comment with its source label (label), new label (claude_label), second choice (claude_also_ok) and whether they agree (agree: yes, also_ok or no). Its split column is the split before step 5.

Margin

Margin is the comment's emotion minus the hardest other emotion, by cosine similarity, on the released rows:

Split	Mean	Median	Wrong (≤ 0)	Easy (≥ 0.10)
train	−0.000	+0.000	49.5%	0.4%
test	−0.002	+0.001	47.7%	0.4%

Nearly every margin lies between −0.10 and +0.10: the model barely separates the emotions in Darija. Disgust is wrong for 87% of train comments.

Zero-shot baseline

Qwen3-VL-Embedding-2B with no training, predicting the most similar emotion, on the 269 test comments before step 5:

Labels	Instructions	Accuracy	Accuracy@3	Macro F1
Relabelled	None	0.379	0.721	0.377
Relabelled	Comment only	0.402	0.848	0.399
Relabelled	Comment and emotion	0.520	0.829	0.526
Source	Comment and emotion	0.491	0.855	0.497

The instructions and emotion wording were chosen by comparing a few wordings on the test split with the source labels, so these scores are slightly optimistic.

Limitations

• The labels are Claude's, not human annotations. Scores on the test split partly measure agreement with Claude; they are not comparable with results reported on the source dataset.
• The test split is small (264 comments, 29 for surprise), so differences of a few points are within noise.
• Emotion in short comments is often ambiguous: 45.6% of comments had a second emotion named as fitting equally well.

Usage

from datasets import load_dataset

ds = load_dataset("FaisalHejary/dz-emotion-embedding")

License and citation

CC BY 4.0, the same as the source. Credit Houdna-khilouf/Dz-Emotion and its paper (IEEE Xplore 11472633) when you use this dataset.