xnliang98 / bart-glc

Code for BART-GLC Training

Code for EACL 2023 paper: Enhancing Dialogue Summarization with Topic-Aware Global- and Local- Level Centrality.

This is an incomplete version, the complete version will be updated after code cleaning.

The runing process:

1. download the pretrained model from https://huggingface.co/uer/bart-base-chinese-cluecorpussmall and placed on the bart-glc/;
2. pip install -r requirements.txt
3. bash run_csds_glc.sh

Some comments:

• We train our model on 4xV100 32G GPUs, you can change the training settings in run_csds_glc.sh.
• The results of our BART-GLC model are placed on the results/
• If you want to get the rouge, bleu, bertscore, and moverscore, you can move in the eval_summ dir.
• The core code of our paper is placed in the src/models.