If you wish to use this model for commercial purposes, please obtain a license by contacting: info@tabularis.ai

🎭 Multilingual Emotion Classification Model (23 Languages, 11 Emotions)

Model Details

Model Name: tabularisai/multilingual-emotion-classification
Base Model: FacebookAI/xlm-roberta-base
Task: Multi-label Text Classification (Emotion Recognition)
Languages: 23 — English, Mandarin Chinese (中文), Spanish (Español), Hindi (हिन्दी), Arabic (العربية), Bengali (বাংলা), Portuguese (Português), Russian (Русский), Japanese (日本語), German (Deutsch), Indonesian (Bahasa Indonesia), Tamil (தமிழ்), Vietnamese (Tiếng Việt), Korean (한국어), French (Français), Turkish (Türkçe), Italian (Italiano), Polish (Polski), Ukrainian (Українська), Urdu (اردو), Dutch (Nederlands), Punjabi (ਪੰਜਾਬੀ), and Swahili.
Number of Classes: 11 — anger, contempt, disgust, fear, frustration, gratitude, joy, love, neutral, sadness, surprise
Label Mode: Multi-label — each text can be assigned zero, one, or multiple emotions (independent sigmoid heads, τ = 0.5).
Usage:
- Social media emotion analysis
- Customer feedback analysis
- Product review emotion tagging
- Brand monitoring
- Conversational AI / chatbot affect tracking
- Market research
- Customer service optimization

Model Description

This model is a fine-tuned version of FacebookAI/xlm-roberta-base for multilingual multi-label emotion classification. It was trained on synthetic multilingual data covering 23 languages and 11 emotion categories, enabling robust emotion detection across languages, registers, and cultural contexts.

Unlike single-label sentiment classifiers, this model predicts a set of emotions per input — reflecting the reality that utterances often carry mixed affect (e.g. gratitude + love, frustration + sadness).

Training Data

Trained on synthetic multilingual data generated by advanced LLMs, providing broad coverage of emotion expressions across all 23 supported languages. All labels are multi-hot vectors over the 11 emotion classes.

Training Procedure

Fine-tuned for 3 epochs with BCEWithLogitsLoss (independent-binary-per-label).
Cosine LR schedule with 6% warmup, lr=2e-5, effective batch size 64.
Mixed precision (bf16) on a single A100; max sequence length 192.
Per-epoch checkpointing; epoch 3 was selected by validation F1-micro.

Evaluation (held-out multilingual test set, 11,500 rows)

Metric	Value
F1 (micro)	0.840
F1 (macro)	0.839
Jaccard (samples)	0.794
Subset accuracy	0.640
Hamming accuracy	0.953
AUROC (micro)	0.980
Average Precision (micro)	0.923
LRAP	0.936

Decision threshold: τ = 0.5 applied independently per label.

Intended Use

Ideal for:

Multilingual social media emotion monitoring
International customer feedback affect analysis
Global product review emotion tagging
Worldwide brand sentiment & emotion tracking
Affect-aware conversational systems

How to Use

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "tabularisai/multilingual-emotion-classification"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
model.eval()

LABELS = ["anger", "contempt", "disgust", "fear", "frustration",
          "gratitude", "joy", "love", "neutral", "sadness", "surprise"]

@torch.no_grad()
def predict_emotions(texts, threshold: float = 0.5):
    inputs = tokenizer(texts, return_tensors="pt", truncation=True,
                       padding=True, max_length=192)
    probs = torch.sigmoid(model(**inputs).logits).cpu().numpy()
    results = []
    for row in probs:
        picked = [(LABELS[i], float(row[i])) for i in range(len(LABELS)) if row[i] >= threshold]
        picked.sort(key=lambda x: -x[1])
        results.append(picked or [("neutral", float(row[LABELS.index("neutral")]))])
    return results


texts = [
    # English
    "Thank you so much for helping me, I really appreciate it!",
    "I can't believe they cancelled the flight again, this is ridiculous.",
    # Spanish
    "¡Qué alegría verte después de tanto tiempo!",
    "Estoy muy decepcionado con el servicio.",
    # Chinese
    "收到你的礼物我真的很感动,谢谢你!",
    "这部电影太吓人了,我都不敢一个人看。",
    # Arabic
    "أنا ممتن جدًا لكل ما فعلته من أجلي.",
    "لا أستطيع تحمّل هذا الوضع أكثر من ذلك.",
    # Hindi
    "आपका यह तोहफ़ा देखकर मेरी आँखों में आँसू आ गए।",
    "यह सेवा बिल्कुल घटिया थी, मैं बहुत निराश हूँ।",
    # Japanese
    "久しぶりに会えて本当に嬉しいです!",
    "また電車が遅れた...本当にうんざりする。",
    # French
    "Je suis tellement reconnaissant pour tout ce que tu as fait.",
    "C'est inadmissible, j'en ai assez de cette situation.",
    # Swahili
    "Asante sana kwa msaada wako, nakupenda sana!",
    "Nimechoka kabisa na huduma hii mbaya.",
]

for t, r in zip(texts, predict_emotions(texts)):
    tags = ", ".join(f"{lbl}({p:.2f})" for lbl, p in r)
    print(f"Text: {t}\nEmotions: {tags}\n")

Using pipelines (returns probability for each of the 11 classes):

from transformers import pipeline

pipe = pipeline(
    "text-classification",
    model="tabularisai/multilingual-emotion-classification",
    function_to_apply="sigmoid",
    top_k=None,
)

print(pipe("I love this product! It's amazing and works perfectly."))

Ethical Considerations

Synthetic training data reduces annotator bias and broadens language coverage, but real-world validation is strongly advised before deploying in high-stakes settings. Emotion labels are culturally situated — predictions should be treated as probabilistic signals, not ground truth about a person's internal state.