Welcome to the official code repository for our EMNLP2024 submission: "Automated Tone Transcription and Clustering with Tone2Vec."
This repository contains all the experimental details and code discussed in our paper.
- Code_emnlp2024.ipynb: Jupyter notebook containing all experimental code, conducted in Google Colab.
- emnlp_weights/: Directory containing pre-calculated Tone2Vec representations.
| Model | Accuracy (%) | Variance |
|---|---|---|
| Best CNN | 61.01 | 0.1052 |
| MLP (3) | 11.94 | 0.2880 |
| MLP (4) | 10.54 | 0.2015 |
| MLP (5) | 12.65 | 0.2089 |
| MLP (6) | 20.84 | 0.1894 |
| MLP (7) | 26.93 | 0.1669 |
| Model | Accuracy (%) | Variance |
|---|---|---|
| Best CNN | 61.01 | 0.1052 |
| XGBoost | 24.12 | 0.2625 |
| SVM | 39.34 | 0.1984 |
| Random Forest | 16.63 | 0.2667 |
| GBM | 24.12 | 0.2115 |
| KNN | 13.11 | 0.2627 |
- VGG: Total trainable parameters: 134,271,683
- ResNet: Total trainable parameters: 11,171,779
- DenseNet: Total trainable parameters: 6,950,659
The notebook is structured to align with the paper's sections and tables:
- Tone2Vec (Section 5): Focus on pitch-based similarity tone representation, dialect clustering, and variance analysis.
- Tone Transcription (Section 6.2): Details methods and results for tone transcription accuracy across different conditions (Tables 2 and 3).
- Tone Clustering (Section 7.2): Results in Tables 4, 5, and Figure 6, exploring tone clustering to identify data patterns and groupings.
- Additional Experiments: Conducted during the rebuttal stage to address reviewer feedback and validate findings.
Please Note: This is an anonymous version intended for review purposes only.
The official package will be released upon acceptance of the paper. Please do not distribute this version.