SenticCrystal

State-of-the-art Emotion Recognition in Conversation with Linguistic Analysis

Achieving 82.69% (4-way) and 67.07% (6-way) weighted F1 on IEMOCAP using strictly causal context (text-only).

Highlights

• SOTA Performance: Outperforms bidirectional models using only causal (past) context
• Statistical Rigor: 10 random seeds, paired t-tests, Bonferroni correction
• Linguistic Analysis: Emotion-specific discourse marker positioning patterns (exploratory)

Key Findings

Finding	Effect	p-value	Status
Context is crucial	+22% F1	<10⁻¹⁵	Confirmed
Hierarchical helps (utterance-only)	+1.9%	0.007	Confirmed
Context subsumes hierarchy	-	0.43	Confirmed
SenticNet doesn't help	0%	-	Negative
Discourse marker patterns	Exploratory	0.01*	Exploratory

Performance Comparison

Method	4-way	6-way	Context Type
HCAM (2023)	81.4%	64.4%	Bidirectional
Mai et al. (2019)	81.5%	-	Intra-utterance
Ours	82.69%	67.07%	Causal

Using strictly causal context (no future utterances) - applicable to real-time systems

Method

• Encoder: Sentence-RoBERTa (nli-roberta-base-v2)
• Context: K-turn sliding window (K=0~200)
• Pooling: Layer averaging (avg_last4) + mean pooling
• Classifier: LSTM with Bayesian hyperparameter optimization (Optuna)

Installation

conda env create -f environment.yml
conda activate senticcrystal

Or with pip:

pip install -r requirements.txt

Project Structure

SenticCrystal/
├── src/                          # Core modules
│   ├── models/                   # Model definitions
│   ├── features/                 # Feature extraction (S-RoBERTa)
│   └── utils/                    # Utilities
│
├── scripts/
│   ├── statistical_analysis/     # Statistical tests (6 scripts)
│   ├── turn/                     # Turn-level training
│   ├── trainer/                  # Utterance-level training
│   └── generator/                # Embedding generation
│
├── results/
│   ├── STATISTICAL_ANALYSIS_SUMMARY.md
│   └── discourse_markers/        # DM position data
│
└── data/                         # IEMOCAP (not included)

Usage

Run Statistical Analysis

python scripts/statistical_analysis/run_all.py

Train Turn-level Model

python scripts/turn/train_turnlevel_k_sweep_bayesian.py \
    --task 4way \
    --encoder sentence-roberta \
    --seed 42

Generate Embeddings

python scripts/generator/generate_sroberta_npz_4way.py
python scripts/generator/generate_sroberta_hier_npz_4way.py

Roadmap

• Text-only baseline (current)
• Multimodal integration (audio features)
• Real-time inference pipeline

Citation

@article{senticcrystal2024,
  title={Understanding Emotion in Discourse: From Recognition to Generation-Informed Insights},
  author={Anonymous},
  journal={TACL (under review)},
  year={2024}
}

License

MIT License

Contact

Questions: cheonkamjeong@gmail.com