Hierarchical Dialogue Understanding with Special Tokens and Turn-level Attention, ICLR 2023, Tiny Papers

https://arxiv.org/abs/2305.00262

This is the official code repository of our ICLR 2023 Tiny paper. In this paper, we proposed a simple but effective Hierarchical Dialogue Understanding model, HiDialog. we first insert multiple special tokens into a dialogue and propose the turn-level attention to learn turn embeddings hierarchically. Then, a heterogeneous graph module is leveraged to polish the learned embeddings.

Our experiments are conducted with following core packages:

• PyTorch 1.11.0
• CUDA 11.6
• dgl-cuda11.3 0.8.2
• sklearn

Main Results

Reproducibility

To reproduce our training process in main experiments on DialogRE,

• download RoBERTa and unzip it to pre-trained_model/RoBERTa/.
• download config.json, merges.txt and vocab.json from here, put them to pre-trained_model/RoBERTa/
• download DialogRE
• copy the *.json files into datasets/DialogRE
• run bash dialogre.sh

To reproduce our training process in main experiments on MELD,

• download RoBERTa and unzip it to pre-trained_model/RoBERTa/.
• download config.json, merges.txt and vocab.json from here, put them to pre-trained_model/RoBERTa/
• download MELD
• copy the *.json files into datasets/MELD
• run python MELD.py
• run bash meld.sh

This project is expanded upon from a course project at NUS [Course Page]. The code repository is based on following projects:

• ACL-20, "Dialogue-Based Relation Extraction" [github]
• AAAI-21, "GDPNet: Refining Refining Latent Multi-View Graph for Relation Extraction" [github]
• EMNLP-21, "Graph Based Network with Contextualized Representations of Turns in Dialogue" [github]

Thanks for their amazing work.

@Article{liu2023hierarchical,
  author  = {Xiao Liu and Jian Zhang and Heng Zhang and Fuzhao Xue and Yang You},
  title   = {Hierarchical Dialogue Understanding with Special Tokens and Turn-level Attention},
  journal = {arXiv preprint arXiv:2305.00262},
  year    = {2023},
}