EmoC

Emotion Intelligence System — real-time emotion recognition and prediction.

Overview

EmoC is an end-to-end emotion intelligence system that classifies text into emotional categories and provides interactive prediction capabilities. The system combines transformer-based classification with a user-facing interactive demo.

Problem

Emotion detection in text is fundamental to understanding human communication. However, existing solutions are either too broad (positive/negative only), too slow for interactive use, or lack interpretability. EmoC addresses all three gaps.

Architecture

`

Input Text → Tokenizer → Transformer Encoder → Classification Head → Emotion Label

↓

Attention Weights

↓

Interpretation Layer

`

Emotion Categories

The system classifies text into 7 distinct emotions based on dimensional affect theory:

Category	Description	Valence-Arousal
----------	-------------	-----------------
Joy	Positive, high-arousal	+1, +1
Sadness	Negative, low-arousal	-1, -1
Anger	Negative, high-arousal	-1, +1
Fear	Negative, low-to-medium arousal	-1, +0.5
Surprise	Neutral, high-arousal	0, +1
Disgust	Negative, medium arousal	-1, 0
Neutral	Low arousal, neutral valence	0, 0

Model Architecture

Base Model: Fine-tuned BERT-base (uncased)

Classification Head: Linear(768, 256) → ReLU → Dropout(0.3) → Linear(256, 7)

Attention Head: Extracts attention weights from layer 8 for interpretability

Token-Level Classification: Average pooling over token embeddings before classification

Training Configuration

Parameter	Value
-----------	-------
Base Model	bert-base-uncased
Optimizer	AdamW (lr=2e-5, weight_decay=0.01)
Scheduler	Linear warmup (10%) + linear decay
Epochs	5
Batch Size	32
Max Sequence Length	128
Dropout	0.3
Hardware	Single NVIDIA T4 (Google Colab)
Training Time	~2.5 hours

Dataset

Primary Dataset: GoEmotions (Google Research)

58,000 Reddit comments

27 emotion categories (mapped to 7 core emotions)

Balanced through oversampling of underrepresented classes

Preprocessing:

Removing HTML and markdown artifacts

Lowercasing (BERT uncased compatibility)

Truncating sequences >128 tokens

Handling class imbalance with weighted sampling

Train/Val/Test split: 80/10/10

Augmentation:

Synonym replacement (WordNet)

Random deletion (p=0.1)

Back-translation for minority classes

Emotion Classification

Classification Head

The classification head is a simple feed-forward network:

`

Pooled Output (768) → Dense(256) → ReLU → Dropout(0.3) → Dense(7) → Softmax

`

Multi-Label Extension

While the primary task is single-label classification, the system also supports multi-label outputs where text can express blended emotions. This is achieved through a secondary sigmoid-based head.

Interpretability

EmoC provides attention-based explanations:

Highlights tokens that most influenced the classification

Shows attention weight distribution across layers

Visualizes which words pushed toward each emotion category

Training

Training Process

Warmup: 500 steps of linear learning rate increase to 2e-5

Fine-tuning: 5 epochs with batch size 32

Gradient clipping: Max norm of 1.0 to prevent exploding gradients

Checkpointing: Save best model based on validation macro-F1

Early stopping: Patience of 2 epochs, monitored on validation loss

Loss Function

Label-smoothed cross-entropy with smoothing parameter ε=0.1. This prevents the model from becoming overconfident and improves generalization to unseen emotional expressions.

Evaluation

Results

Metric	Value
--------	-------
Accuracy	72.4%
Macro F1	68.1%
Weighted F1	71.8%
Per-Class F1 (avg)	66.3%
Confusion Matrix	See below

Per-Class Performance

Emotion	Precision	Recall	F1
---------	-----------	--------	-----
Joy	0.78	0.71	0.74
Sadness	0.69	0.74	0.71
Anger	0.73	0.68	0.70
Fear	0.64	0.62	0.63
Surprise	0.71	0.75	0.73
Disgust	0.68	0.66	0.67
Neutral	0.74	0.73	0.73

Error Analysis

Joy ↔ Surprise: Confused when surprise is positive and high-arousal

Anger ↔ Disgust: Overlapping negative high-arousal features

Fear ↔ Sadness: Both low-valence emotions with subtle arousal differences

Applications

Customer service: Detect customer frustration in real-time

Mental health: Monitor emotional state in therapy chatbots

Content moderation: Flag emotionally charged content

Education: Adapt content difficulty based on student emotional state

Market research: Analyze sentiment in product reviews at the emotion level

Interactive Prediction Demo

The system includes a live demo where users can:

Type or paste text

See real-time emotion classification with confidence scores

View attention visualization highlighting influential words

Compare predictions across emotion categories

Submit feedback to improve model performance

The demo is powered by a fast inference server with cached model embeddings for sub-100ms response times.

Lessons Learned

BERT is a strong baseline: Fine-tuning BERT-base for emotion classification is surprisingly effective without custom architectures.

GoEmotions is noisy: Human annotation of emotions has inherent ambiguity — don't expect >80% accuracy on fine-grained categories.

Attention is not always trustworthy: Layer 8 attention weights correlate well with final predictions but can be misleading for some examples.

Class imbalance is critical: Weighted sampling improved macro F1 by 8% over uniform sampling.

Interpretability drives trust: Users prefer models that show attention highlights over black-box predictions.

Future Work

Extend to multilingual emotion recognition

Add emotion dynamics tracking across conversations

Integrate visual cues (facial expression) for multimodal emotion detection

Explore emotion causality — what causes specific emotional shifts in text

Build a real-time API with WebSocket streaming support

Source Code

GitHub Repository