This is the official repository for the COLING 2025 paper: "Non-Emotion-Centric Empathetic Dialogue Generation".
本仓库是 COLING 2025 论文 "Non-Emotion-Centric Empathetic Dialogue Generation" 的官方代码。
├── main.py # Training entrypoint (supports DDP multi-GPU) / 训练入口（支持DDP多卡）
├── train.py # Training & validation logic / 训练和验证逻辑
├── run_comet.py # COMET commonsense knowledge generation / COMET常识知识生成
├── train_entity_ranker.py # Entity ranking model training / 实体排序模型训练
├── constant.py # Emotion labels & relation mappings / 情感标签和关系映射
├── config.yaml # Configuration file / 配置文件
├── model/
│ ├── model.py # CustomBartForConditionalGeneration & BertForRanking / 模型定义
│ └── Decoder.py # Custom BART decoder with knowledge attention / 带知识注意力的自定义解码器
├── data/
│ ├── util.py # Beam search, metrics, utilities / Beam search、评估指标、工具函数
│ └── datasets/
│ ├── empchat.py # EmpDataset & RankingDataset / 数据集类
│ └── loader.py # DataLoader collate functions / 数据加载器
└── metrics/
└── distinct/
└── distinct.py # Distinct-N metric / Distinct-N 指标
# Create virtual environment / 创建虚拟环境
python -m venv .venv
source .venv/bin/activate
# Install dependencies / 安装依赖
pip install -r requirements.txt
# Download NLTK data / 下载NLTK数据
python -c "import nltk; nltk.download('punkt'); nltk.download('punkt_tab'); nltk.download('averaged_perceptron_tagger'); nltk.download('stopwords')"Download the following models and place them under pretrained/:
下载以下模型并放置到 pretrained/ 目录下：
| Model / 模型 | Source / 来源 | Path / 路径 |
|---|---|---|
| BART-base | facebook/bart-base | pretrained/bart-base/ |
| MPNet-base | sentence-transformers/all-mpnet-base-v2 | pretrained/mpnet-base/ |
| BERT-base | bert-base-uncased | pretrained/bert-base/ |
| COMET-ATOMIC 2020 | COMET-ATOMIC_2020_BART | pretrained/comet-atomic_2020_BART/ |
You can skip Steps 1–4 entirely by downloading the pre-built dataset, which includes raw data, COMET knowledge, the trained entity ranking model, and dataset caches.
你可以通过下载预先构建好的数据集来跳过下面的 Step 1–4，该数据集已包含原始数据、COMET 知识、训练好的实体排序模型和数据集缓存。
Download link / 下载链接: [TODO: add link]
After downloading, place the files as follows / 下载后按如下结构放置：
├── DCKS-dataset/
│ ├── train.pkl # Raw data / 原始数据
│ ├── val.pkl
│ ├── test.pkl
│ ├── DCKS-all_train_comet_social_pickle.pkl # COMET social knowledge / COMET社交知识
│ ├── DCKS-all_test_comet_social_pickle.pkl
│ ├── DCKS-all_train_comet_entity_pickle.pkl # COMET entity knowledge / COMET实体知识
│ ├── DCKS-all_test_comet_entity_pickle.pkl
│ ├── DCKS-train_dataset.json # Dataset cache / 数据集缓存
│ └── DCKS-test_dataset.json
└── DCKS-entity_ranking_model_context_best/ # Trained entity ranking model / 训练好的实体排序模型
├── config.json
├── pytorch_model.bin
├── tokenizer.json
├── tokenizer_config.json
├── special_tokens_map.json
└── vocab.txt
Once placed, you can go directly to Training / 训练.
放置完成后可直接跳到 训练 部分。
If you prefer to build the dataset from scratch, follow Steps 1–4 below.
如果你希望从头构建数据集，请按以下 Step 1–4 操作。
Place the EmpatheticDialogues dataset pickle files (train.pkl, val.pkl, test.pkl) under DCKS-dataset/.
将 EmpatheticDialogues 数据集的 pickle 文件（train.pkl、val.pkl、test.pkl）放到 DCKS-dataset/ 目录下。
python run_comet.py --config_path config.yamlThis generates social and entity commonsense knowledge for each dialogue example.
为每条对话样本生成社交和实体常识知识。
python train_entity_ranker.py --cuda-id 0 --config-path config.yamlThe first run of training will automatically build and cache the processed dataset (JSON files under DCKS-dataset/).
首次训练时会自动构建并缓存处理后的数据集（JSON 文件保存在 DCKS-dataset/ 下）。
Set gpu_ids: [0] in config.yaml, then:
在 config.yaml 中设置 gpu_ids: [0]，然后：
python main.py --config_path config.yamlSet gpu_ids to the list of GPU IDs you want to use:
将 gpu_ids 设置为要使用的 GPU ID 列表：
runtime:
gpu_ids: [0, 1, 2]python main.py --config_path config.yamlThe training script uses torch.multiprocessing.spawn internally — no need for torchrun.
训练脚本内部使用 torch.multiprocessing.spawn，无需使用 torchrun。
| Key | Description / 说明 | Default |
|---|---|---|
data_folder |
Dataset directory / 数据集目录 | DCKS-dataset |
model_save_folder |
Model save directory / 模型保存目录 | model_save |
bart_model_path |
BART model path / BART模型路径 | pretrained/bart-base |
mpnet_model_path |
MPNet model path / MPNet模型路径 | pretrained/mpnet-base |
device |
Default device / 默认设备 | cuda:0 |
| Key | Description / 说明 | Default |
|---|---|---|
gpu_ids |
GPU IDs for DDP / DDP使用的GPU列表 | [0] |
epoch |
Number of epochs / 训练轮数 | 10 |
per_gpu_batch_size |
Batch size per GPU / 每张卡的batch size | 4 |
wo_entity |
Disable entity knowledge / 关闭实体知识 | false |
wo_social |
Disable social knowledge / 关闭社交知识 | false |
| Key | Description / 说明 | Default | Recommended Range / 推荐范围 |
|---|---|---|---|
CL |
Enable contrastive learning / 启用对比学习 | true |
true/false |
CL_sample_num |
Negative samples per type / 每种负样本数量 | 3 |
2, 3, 5 |
alpha_for_CL |
Weight for semantic similarity vs LM score / 语义相似度与LM分数的权重 | 0.7 |
0.3, 0.5, 0.7, 0.9 |
self_generated |
Use model-generated negatives / 使用模型自生成负样本 | true |
true/false |
emotion_nega |
Use emotion-based negatives / 使用情感负样本 | true |
true/false |
high_freq_nega |
Use high-frequency negatives / 使用高频句负样本 | true |
true/false |
train_beam_size_for_CL |
Beam size for self-generated negatives / 自生成负样本的beam size | 10 |
5, 10, 15 |
cl_max_candidates |
Max CL candidates per sample / 每个样本最大CL候选数 | 64 |
32, 64, 128 |
cl_ranking_margin |
Base margin for ranking loss / 排序损失基础margin | 0.01 |
0.005, 0.01, 0.05 |
cl_gold_bleu_threshold |
BLEU threshold to mask gold / 过滤金标准的BLEU阈值 | 0.99 |
0.95, 0.99 |
cl_bleu_ngram |
N-gram order for CL BLEU / CL中BLEU的n-gram阶数 | 2 |
2, 3 |
The contrastive learning module supports three types of negative samples:
对比学习模块支持三种负样本：
-
Batch-internal negatives (always on when CL=true): Other responses in the same batch serve as negatives. / Batch内负样本（CL=true时始终开启）：同一batch中的其他回复作为负样本。
-
Emotion-based negatives (
emotion_nega): Responses from different emotion categories. Controlled byCL_sample_num. / 情感负样本（emotion_nega）：来自不同情感类别的回复，数量由CL_sample_num控制。 -
Self-generated negatives (
self_generated): Model's own beam search outputs. Most expensive but most effective. Beam size controlled bytrain_beam_size_for_CL. / 自生成负样本（self_generated）：模型自身beam search的输出。开销最大但效果最好，beam size 由train_beam_size_for_CL控制。 -
High-frequency negatives (
high_freq_nega): Most common sub-sentences from training data. Low overhead. / 高频句负样本（high_freq_nega）：训练数据中最常见的子句。开销很低。
Tips / 建议:
- Start with
emotion_negaonly for fast iteration, then addself_generatedfor best results. / 先只开emotion_nega快速迭代，再加self_generated获得最佳效果。 self_generatedsignificantly increases training time (~3x). Use multi-GPU DDP to compensate. /self_generated会显著增加训练时间（约3倍），建议使用多卡DDP加速。cl_max_candidates> 64 may cause OOM on 24GB GPUs with batch_size=4. /cl_max_candidates> 64 在24GB显存、batch_size=4时可能OOM。
The model is evaluated on:
- BLEU-1/2/3/4: N-gram overlap with reference responses / 与参考回复的N-gram重叠度
- Distinct-1/2/3: Generation diversity / 生成多样性
- Perplexity: Language model quality / 语言模型质量
Validation runs automatically after each epoch (starting from epoch 2). The best model (by BLEU-4) is saved to model_save/best_model/.
验证在每个 epoch 后自动运行（从第2个epoch开始）。最佳模型（按BLEU-4）保存到 model_save/best_model/。
@inproceedings{huang2025non,
title={Non-Emotion-Centric Empathetic Dialogue Generation},
author={Huang, Yuanxiang},
booktitle={Proceedings of the 31st International Conference on Computational Linguistics (COLING 2025)},
year={2025}
}This project is licensed under the Apache License 2.0. See LICENSE for details.
本项目采用 Apache License 2.0 许可证。详见 LICENSE。