Sentimotion_v1

Model Details

Model Name

Sentimotion_v1

Model Type

Joint Sentiment Analysis and Emotion Classification model for Bangla (Bengali) text.

Base Model

csebuetnlp/banglabert_large

Architecture

• BanglaBERT-large encoder
• Shared contextual representation (CLS token)
• Two parallel classification heads:
  • Sentiment head (3 classes)
  • Emotion head (6 classes)

Tasks

• Sentiment Classification
• Emotion Classification

Language(s)

• Bangla (bn)

---

Intended Use

Primary Use Cases

This model is intended for:

• Bangla social media sentiment analysis
• Political opinion and discourse monitoring
• Emotion-aware text analytics
• Public opinion mining
• Academic and applied Bangla NLP research
• Human-in-the-loop moderation and analytics systems

Out-of-Scope Uses

This model should not be used:

• As the sole system for content moderation
• For legal, medical, or law-enforcement decisions
• For automated decision-making affecting individuals

Human review is strongly recommended for sensitive or high-impact applications.

---

Model Inputs and Outputs

Input

A single Bangla text string.

Example: জনগণ নির্ধারণ করবে কে জয়ী হবে।

import torch
import numpy as np

SENTIMENT_LABELS = ["negative", "neutral", "positive"]
EMOTION_LABELS = ["neutral", "happy", "anger_disgust", "sadness", "surprise", "fear"]

def predict(text, model, tokenizer, device="cpu",
            sent_threshold=0.55,
            emo_threshold=0.60):

    model.eval()

    inputs = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        max_length=128
    ).to(device)

    with torch.no_grad():
        sent_logits, emo_logits = model(
            inputs["input_ids"],
            inputs["attention_mask"]
        )

    # ---------- Softmax ----------
    sent_probs = torch.softmax(sent_logits, dim=1).cpu().numpy()[0]
    emo_probs  = torch.softmax(emo_logits,  dim=1).cpu().numpy()[0]

    # ---------- Sentiment ----------
    sent_idx = int(np.argmax(sent_probs))
    sent_conf = float(sent_probs[sent_idx])

    if sent_conf < sent_threshold:
        sentiment = "neutral"
        sent_conf = float(sent_probs[SENTIMENT_LABELS.index("neutral")])
    else:
        sentiment = SENTIMENT_LABELS[sent_idx]

    # ---------- Emotion ----------
    emo_idx = int(np.argmax(emo_probs))
    emo_conf = float(emo_probs[emo_idx])

    if emo_conf < emo_threshold:
        emotion = "neutral"
        emo_conf = float(emo_probs[EMOTION_LABELS.index("neutral")])
    else:
        emotion = EMOTION_LABELS[emo_idx]

    # ---------- JSON output ----------
    result = {
        "sentiment": sentiment,
        "sentiment_confidence": round(sent_conf, 6),
        "emotion": emotion,
        "emotion_confidence": round(emo_conf, 6),
        "sentiment_scores": {
            label: float(sent_probs[i])
            for i, label in enumerate(SENTIMENT_LABELS)
        },
        "emotion_scores": {
            label: float(emo_probs[i])
            for i, label in enumerate(EMOTION_LABELS)
        }
    }

    return result

Maximum supported length:

• Up to 512 tokens (recommended: 128 tokens)

---

Output

The model produces two independent predictions:

Sentiment (3-class)

• negative
• neutral
• positive

Emotion (6-class)

• neutral
• happy
• anger_disgust
• sadness
• surprise
• fear

Each output is returned as a softmax probability distribution.

---

Training Details

Training Strategy

• Multi-task learning with a shared encoder
• Joint optimization of sentiment and emotion heads
• Balanced cross-entropy loss between tasks

Dataset

• Custom labeled Bangla dataset
• Includes real-world Bangla text such as:
  • political commentary
  • public opinion
  • social media–style language

Data Characteristics

• Sentiment labels consolidated into 3 classes
• Emotion labels are single-label (dominant emotion)
• Strong class imbalance for fear and surprise

Hardware

• GPU-based training
• Mixed precision (FP16)

---

Evaluation Results

Sentiment Classification (Test Set)

Metric	Score
Accuracy	~0.87
Macro F1	~0.86

Class	F1-score
Negative	~0.90
Neutral	~0.82
Positive	~0.86

---

Emotion Classification (Test Set)

Metric	Score
Macro F1	~0.63
Weighted F1	~0.77

Emotion	F1-score	Notes
Neutral	~0.73	Strong
Happy	~0.83	Strong
Anger/Disgust	~0.84	Very strong
Sadness	~0.67	Moderate
Surprise	~0.26	Low (few samples)
Fear	~0.44	Low (few samples)

Lower performance on fear and surprise is primarily due to limited labeled data, not architectural limitations.

---

Ethical Considerations

Biases

• Political text may trigger higher sensitivity to anger-related language
• Predictions reflect the distribution and biases of the training data

Mitigation Strategies

• Use probability thresholds instead of hard labels
• Combine with a separate hate-speech or abuse detection model
• Maintain human oversight for moderation or policy use

---

Limitations

• Not a hate-speech detector by itself
• Rare emotion classes generalize less effectively
• Emotion predictions capture dominant signals, not mixed emotions
• Attention/gradient-based explanations indicate influence, not causality

---

How to Use the Model

⚠️ This model uses a custom architecture and cannot be loaded directly using
AutoModelForSequenceClassification.

Users must:

1. Load the base BanglaBERT encoder
2. Reconstruct the joint sentiment–emotion architecture
3. Load weights from model.safetensors

Refer to the repository for example inference code and API usage.

---

License

Apache License 2.0

---

Citation

If you use this model in academic or applied research, please cite:

@misc{sentimotion2026,
  title  = {Sentimotion_v1: A Joint Bangla Sentiment and Emotion Classification Model},
  author = {Arafat Fahim},
  year   = {2026},
  url    = {https://huggingface.co/arafatfahim/sentimotion}
}