Single-Pass, Depth-Selective Reading for Multi-Aspect Sentiment Analysis (Official Site)

This repository contains the official code for our ACL 2026 (main, long paper) - Single-Pass, Depth-Selective Reading for Multi-Aspect Sentiment Analysis

Updated on June 18, 2026 (Paper is uploaded)

Released on April 13, 2026

---

🚀 Introduction

DABS is a single-pass inference framework for Aspect-Term Sentiment Analysis in multi-aspect sentences. It encodes each sentence once to construct a reusable, depth-ordered substrate, and then performs aspect-conditioned readout without re-encoding. The framework consists of:

• DORA, which constructs a shared depth substrate via a single encoder pass.
• ACBS, which performs aspect-conditioned token localization and budget-aware depth selection.

Experiments on four ATSA benchmarks show that DABS achieves competitive performance while reducing end-to-end computation by up to 60% in multi-aspect settings (M > 2).

Repository Layout

.
├── data/                        # benchmark data
│   └── semeval/                 # SemEval ATSA benchmarks and multilingual variants
├── figures/                     # paper figures
│   └── DABS_framework.png       # framework overview
├── outputs/                     # checkpoints and evaluation outputs
├── results/                     # batch logs and summaries
├── scripts/                     # runnable entry points
│   ├── train.py                 # single-run training and evaluation
│   ├── run_full_model_batch.py  # full-model batch runs
│   └── run_multilingual_ablation.py  # multilingual experiments
├── src/                         # source code
│   ├── config/                  # configuration and dataset recipes
│   ├── core/                    # DABS, DORA, and ACBS implementation
│   └── utils/                   # callbacks, saving, and preprocessing utilities
├── README.md                    # English README
├── README-CHINESE.md            # Chinese README
└── requirements.txt             # package versions

🛠️ Requirements

Install dependencies with:

pip install -r requirements.txt

Notes:

• Because this project was developed on an RTX 5090 GPU, some packages were early locally built versions. If exact builds are unavailable, please use approximately matching versions.
• Do not use transformers 5.x. Parts of the current codebase are not compatible with it.

Unified Data Loading

We have integrated a unified entry point for all SemEval datasets in src/core/data.py, and we encourage other projects to directly reuse this loader for data access and preprocessing. The repository includes the main English SemEval ATSA benchmarks together with multilingual Restaurant-16 variants:

• 2: Laptop-14
• 3: Restaurant-14
• 4: Restaurant-15
• 5: Restaurant-16
• 6: Restaurant-16-FR
• 7: Restaurant-16-RU
• 8: Restaurant-16-ES
• 9: Restaurant-16-DU
• 10: Restaurant-16-TU

The repository expects the processed JSON files under:

• data/semeval/Laptop_14/
• data/semeval/Restaurant_14/
• data/semeval/Restaurant_15/
• data/semeval/Restaurant_16/
• data/semeval/Restaurant_16_FR/
• data/semeval/Restaurant_16_RU/
• data/semeval/Restaurant_16_ES/
• data/semeval/Restaurant_16_DU/
• data/semeval/Restaurant_16_TU/

Download Data from Google Drive

The data/ directory is distributed as a compressed archive instead of being maintained directly in the GitHub repository.

• Dataset archive (data.tar.gz): Google Drive

After downloading data.tar.gz, place it at the project root and extract it with:

tar -xzf data.tar.gz

This restores the expected data/ directory used by src/core/data.py.

Download Released Full-Model Checkpoints

The released full-model checkpoints are also provided as a compressed archive:

• Checkpoint archive (full_model.tar.gz): Google Drive

After downloading full_model.tar.gz, place it at the project root and extract it with:

tar -xzf full_model.tar.gz

This restores outputs/full_model/. You can then run batch inference over the released checkpoints with:

python scripts/run_full_model_inference.py --device cuda:0

Run DABS

Single run:

DATASET_CHOICE=3 RANDOM_SEED=42 python scripts/train.py --dual-layer

Batch runs for the full model on the four benchmarks:

python scripts/run_full_model_batch.py --datasets 2 3 4 5 --seeds 42 123 456

Outputs are written to:

• outputs/full_model/...
• results/full_model_batch_<timestamp>/...

Reuse vs Non-Reuse Comparison

To compare standard aspect-wise evaluation against the reuse path on a given checkpoint, run:

python scripts/compare_reuse_non_reuse_eval.py outputs/full_model/Restaurant-14/seed_42 --dataset-choice 3

You can also specify a JSON output path if you want to save the comparison report:

python scripts/compare_reuse_non_reuse_eval.py \
  outputs/full_model/Restaurant-14/seed_42 \
  --dataset-choice 3 \
  --json results/reuse_vs_non_reuse_res16_seed42.json

Evaluation Protocol

These benchmarks do not provide a standard development split. Following the protocol in the paper, the best checkpoint within the training budget is selected on the test split by macro-F1.

Citation

If you find our code useful, feel free to ⭐ star this repo!

If you use our work in your research, please cite:

@inproceedings{XiaACL2026,
  title={Single-Pass, Depth-Selective Reading for Multi-Aspect Sentiment Analysis},
  author={Xia, Yan and Pan, Zhuangzhuang and Kamsim, Amirrudin and Chan, Chee Seng},
  booktitle={Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL)},
  year={2026}
}