MominaMahmood/All-Emotion-Adapters

All Emotion Adapters: MAD-X adapters for multilingual emotion detection

Trained MAD-X adapters from my MSc dissertation, "Multilingual Models Have Feelings Too?" (MSc Applied Artificial Intelligence, University of Huddersfield, 2026).

They detect six emotions in text as a multi-label task (a text can have several emotions, or none):

anger, disgust, fear, joy, sadness, surprise

They cover 9 languages across three resource levels, from the BRIGHTER dataset (SemEval-2025 Task 11):

Tier	Languages (code)
High resource	English (eng), Hindi (hin), Russian (rus)
Mid resource	Hausa (hau), Kinyarwanda (kin), Sundanese (sun)
Low resource	Yoruba (yor), Emakhuwa (vmw), Nigerian Pidgin (pcm)

• Dissertation code and results: https://github.com/Momina0398/EmotionDetection
• API that serves these adapters (FastAPI, Docker, CI): https://github.com/Momina0398/emotion-api

What is in this repo

adapters_xlmr_weighted/     47 runs on xlm-roberta-large (all 9 languages)
adapters_serengeti/         26 runs on UBC-NLP/serengeti-E250 (5 African languages)
lang_sun_lapt/              Sundanese language adapter, extra pre-training (LAPT)
lang_vmw_lapt/              Emakhuwa language adapter, LAPT, xlm-roberta-large
lang_vmw_lapt_serengeti/    Emakhuwa language adapter, LAPT, SERENGETI

Each run folder is named <lang>_track_<a|c>_<C1|C2|C3> and contains three parts:

lang/   language adapter (invertible MAD-X language adapter)
task/   emotion task adapter ("emotion_task")
head/   6-label multi-label classification head

Track

• track_a: target-language training data was available (supervised)
• track_c: no target-language training data (cross-lingual transfer only)

Source configuration (which languages the task adapter was trained on)

• C1: all other BRIGHTER languages
• C2: languages in the same top-level family as the target
• C3: languages in the same sub-family as the target

Some combinations are missing on purpose. When a family or sub-family pool was too small to form a valid training set, that run was not created rather than filled with an invented pool.

How to use

# pip install adapters==1.3.0 transformers torch huggingface_hub
import torch
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer
from adapters import AutoAdapterModel
import adapters.composition as ac

REPO = "MominaMahmood/All-Emotion-Adapters"
RUN = "adapters_xlmr_weighted/rus_track_a_C1"
BASE = "xlm-roberta-large"            # use "UBC-NLP/serengeti-E250" for adapters_serengeti runs
LABELS = ["anger", "disgust", "fear", "joy", "sadness", "surprise"]

path = snapshot_download(REPO, allow_patterns=[f"{RUN}/*"])
run = f"{path}/{RUN}"

tok = AutoTokenizer.from_pretrained(BASE)
model = AutoAdapterModel.from_pretrained(BASE)
model.load_adapter(f"{run}/lang", load_as="lang")
model.load_adapter(f"{run}/task", load_as="emotion_task")
model.load_head(f"{run}/head")
model.set_active_adapters(ac.Stack("lang", "emotion_task"))
model.eval()

texts = ["Я так рада тебя видеть!"]
enc = tok(texts, padding=True, truncation=True, max_length=256, return_tensors="pt")
with torch.no_grad():
    probs = torch.sigmoid(model(**enc).logits)[0].tolist()
print({e: round(p, 3) for e, p in zip(LABELS, probs)})
print("labels:", [e for e, p in zip(LABELS, probs) if p >= 0.5])

Use a threshold of 0.5 per label, as in training and evaluation.

Training details

Setting	Value
Architecture	MAD-X: language adapter + task adapter stacked on a frozen backbone
Backbones	xlm-roberta-large, UBC-NLP/serengeti-E250
Adapter configs	SeqBnInvConfig (language), SeqBnConfig (task), reduction factor 16
Loss	Binary cross-entropy with per-label positive class weights ("weighted")
Learning rate	5e-5
Batch size	8
Max epochs	10, early stopping with patience 3 on the dev split
Max sequence length	256
Library	adapters 1.3.0
LAPT	Extra masked-language-model training of the language adapter for Emakhuwa (Emakhuwa text) and Sundanese (Javanese and Indonesian text)

Model selection used the validation split only. Full results, significance tests and per-language analysis are in the dissertation repo linked above.

Limitations

• English has no disgust labels in the BRIGHTER training data, so English runs cannot learn that emotion. Treat English disgust scores as meaningless.
• Single random seed. Results do not capture variation between training runs.
• Emotion labels are dataset annotations, not ground truth about how a writer feels.
• Performance on the low-resource languages (especially Emakhuwa and Nigerian Pidgin) is well below the high-resource languages.
• A main finding of the dissertation is that sub-family transfer (C3) did not beat broader source pools. For most languages, pick C1 or C2 first.
• Not for high-stakes decisions about individuals (for example health, hiring or policing) without expert human review.

Citation

If you use these adapters, please cite the dissertation:

Siddiq, M. M. (2026). Multilingual Models Have Feelings Too? Culturally Aware
Sub-Family Transfer for Emotion Detection. MSc dissertation, University of Huddersfield.

Please also cite the BRIGHTER dataset and SemEval-2025 Task 11 papers.