DyFuLM

KO | EN

gitlite — search

#javascript #python #ai #typescript #go #deep-learning #linux #docker #nodejs #java #vue #awesome-list #ios #openai

(★ 12)

No description available.

DyFuLM Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# [ICIC 2026] DyFuLM: An Advanced Multimodal Framework for Sentiment Analysis

Official code release reconstructed from the authors' experimental notebooks for **dynamic fusion, hierarchical sentiment classification, and intensity regression**.

[English](README.md) | [简体中文](README_zh-CN.md)

[📄 Preprint](https://arxiv.org/abs/2512.01410) · [🔗 DOI](https://doi.org/10.1007/978-981-92-3435-6_43) · [💻 Code](https://github.com/DurianLoop/DyFuLM) · [🧩 Reproduction](#reproduction) · [📝 Citation](#citation)

![Paper](https://img.shields.io/badge/Paper-Springer-B31B1B)
![DOI](https://img.shields.io/badge/DOI-10.1007%2F978--981--92--3435--6__43-2563EB)
![Conference](https://img.shields.io/badge/Conference-ICIC%202026-4F46E5)
![Python](https://img.shields.io/badge/Python-3.10%2B-3776AB)
![Framework](https://img.shields.io/badge/Framework-PyTorch-EE4C2C)
![License](https://img.shields.io/badge/License-MIT-EAB308)

**Ruohan Zhou · Jiachen Yuan · Churui Yang · Wenzheng Huang · Guoyan Zhang · Shiyao Wei · Jiazhen Hu · Ning Xin · Md Maruf Hasan**

</div>

---

## Publication

| Item | Details |
|---|---|
| Title | *DyFuLM: An Advanced Multimodal Framework for Sentiment Analysis* |
| Venue | 22nd International Conference on Intelligent Computing (**ICIC 2026**) |
| Proceedings | *Advanced Intelligent Computing Technology and Applications*, Part XXIV |
| Series | Lecture Notes in Computer Science, volume 16665 |
| Publisher | Springer Nature Singapore |
| Bibliographic year | 2027 |
| DOI | [10.1007/978-981-92-3435-6_43](https://doi.org/10.1007/978-981-92-3435-6_43) |
| Citation files | [BibTeX](citations/paper.bib) · [RIS](citations/paper.ris) · [EndNote](citations/paper.enw) |

> The conference is referred to as ICIC 2026, while the Springer proceedings metadata uses the bibliographic year 2027.

## Overview

DyFuLM treats pretrained language encoders as complementary semantic views. It dynamically selects information from multiple Transformer layers, performs gated cross-model interaction, and jointly learns coarse sentiment, fine sentiment, and normalized sentiment intensity.

The public release is reconstructed from three supplied notebooks. [`END2.ipynb`](docs/notebook_selection.md) was selected as the canonical implementation because its dynamic layer fusion, gated aggregation, hierarchical task guidance, and uncertainty-based multi-task loss align most closely with the paper.

## Architecture

The architecture follows a **RoBERTa–DeBERTa dual-encoder backbone**, hierarchical dynamic fusion, gated feature aggregation, and three shared multi-task prediction heads for coarse-grained sentiment, fine-grained emotion, and sentiment-intensity regression.

## Core Modules

- **Layer-wise dynamic feature extraction:** a BiLSTM and attention mechanism aggregate the final Transformer hidden layers for every token.
- **Hierarchical dynamic fusion:** stacked bidirectional cross-attention layers regulate information exchange between encoder views.
- **Gated feature aggregation:** learned gates balance original and cross-model contextual features.
- **Hierarchical task guidance:** coarse-level sentence information recalibrates sequence features before fine-grained prediction.
- **Dynamic multi-task loss:** learnable uncertainty terms balance coarse classification, fine classification, and intensity regression.

## Notebook-to-Repository Reconstruction

```text
END2.ipynb
├── preprocessing        → src/dyfulm/data/preprocessing.py
├── tokenized dataset    → src/dyfulm/data/dataset.py
├── dynamic layer fusion → src/dyfulm/models/layerwise.py
├── gated fusion         → src/dyfulm/models/fusion.py
├── task heads/model     → src/dyfulm/models/dyfulm.py
├── dynamic loss         → src/dyfulm/losses.py
└── train/evaluate       → src/dyfulm/engine/ + CLI entry points
```

The other two notebooks are documented as experimental baselines in [`docs/notebook_selection.md`](docs/notebook_selection.md).

## Reproduction

### 1. Clone and install

```bash
git clone https://github.com/DurianLoop/DyFuLM.git
cd DyFuLM
conda env create -f environment.yml
conda activate dyfulm
pip install -e .
```

### 2. Prepare the dataset

Place the Booking.com review CSV at:

```text
data/Hotel_Reviews.csv
```

Check the schema:

```bash
python tools/check_dataset.py --csv data/Hotel_Reviews.csv
```

The CSV must contain `Positive_Review`, `Negative_Review`, and `Reviewer_Score`. The dataset itself is not redistributed.

### 3. Train

```bash
python train.py --config configs/default.yaml --prepare-data
```

### 4. Evaluate

```bash
python evaluate.py   --config configs/default.yaml   --checkpoint outputs/default/best.pt   --split test
```

### 5. Predict one review

```bash
python predict.py   --config configs/default.yaml   --checkpoint outputs/default/best.pt   --text "The room was clean and the staff were extremely helpful."
```

More details are available in [`docs/reproduction.md`](docs/reproduction.md).

## Reference Results

The paper reports the following values. They are included for reference and were **not re-run during repository restructuring**.

| Metric | Reported value |
|---|---:|
| Coarse accuracy | 0.8264 |
| Fine accuracy | 0.6848 |
| Coarse F1 | 0.8215 |
| Fine F1 | 0.6814 |
| MAE | 0.0674 |
| MSE | 0.0082 |
| R² | 0.6903 |

## Repository Structure

```text
DyFuLM/
├── assets/                   # Architecture figure
├── citations/                # BibTeX, RIS, and EndNote
├── configs/                  # YAML experiment configuration
├── data/                     # Local data placeholder only
├── checkpoints/              # Checkpoint documentation only
├── docs/                     # Dataset, migration, and reproduction notes
├── scripts/                  # One-command shell entry points
├── src/dyfulm/
│   ├── data/                 # Preparation and tokenized dataset
│   ├── engine/               # Trainer and metrics
│   ├── models/               # Layer fusion, gated fusion, heads, DyFuLM
│   ├── utils/                # Seeds and checkpoints
│   ├── config.py
│   └── losses.py
├── tests/                    # Offline unit tests for custom modules
├── tools/                    # Dataset validation
├── train.py
├── evaluate.py
└── predict.py
```

## Data and Checkpoints

This is a **code-only release**. It does not include the Booking.com dataset, pretrained model caches, fine-tuned checkpoints, or experiment logs. Hugging Face model weights are downloaded on first use unless local paths are supplied in the YAML configuration.

## Citation

```bibtex
@inproceedings{zhou2027dyfulm,
  author    = {Zhou, Ruohan and Yuan, Jiachen and Yang, Churui and Huang, Wenzheng and Zhang, Guoyan and Wei, Shiyao and Hu, Jiazhen and Xin, Ning and Hasan, Md Maruf},
  title     = {DyFuLM: An Advanced Multimodal Framework for Sentiment Analysis},
  booktitle = {Advanced Intelligent Computing Technology and Applications},
  series    = {Lecture Notes in Computer Science},
  volume    = {16665},
  publisher = {Springer Nature Singapore},
  year      = {2027},
  doi       = {10.1007/978-981-92-3435-6_43}
}
```

## Acknowledgements

This implementation uses PyTorch, Hugging Face Transformers, pandas, and scikit-learn.

## License

The code is released under the [MIT License](LICENSE). Datasets and pretrained model assets remain subject to their original terms.

DyFuLM

DyFuLM Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

DyFuLM

DyFuLM Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README