The repository of our work "Topic-Oriented Dialogue Summarization". The paper is under reviewing.
We propose a new dialogue summarization task named Topic-Oriented Dialogue Summarization (TODS). Given a dialogue and a specific topic, the task aims to generate a summary that covers the main content related to the given topic. To achieve this task, we propose three topic-related auxiliary tasks for TODS, including Topic Identification Task, Attention Restriction Task, and Topic Summary Distinguishing Task.
We experiment on two datasets (CSDS and DialogSum) by modifying them to adapt for TODS, and two baseline methods (BART-base and BART-large).
- CSDS dataset: We put the processed CSDS-topic in data/CSDS/topic/.
- DialogSum dataset: We put the processed DialogSum-topic in data/DialogSum/topic/.
- Pretrained BART model:
- python == 3.8
- pytorch == 1.8.1
- files2rouge == 2.1.0
- jieba == 0.42.1
- numpy == 1.21.2
- cytoolz == 0.11.2
- nltk == 3.6.5
- bert-score == 0.3.10
- moverscore == 1.0.3
- transformers == 4.4.1
- Go to the models/ directory.
- Run the bash file run_CSDS.sh to train and test on CSDS-topic
- Run the bash file run_DialogSum.sh to train and test on DIALOGSUM-topic
The reference codes of the provided methods come from:
We thanks for all these researchers who have made their codes publicly available.
We will update the citation format after the paper is accepted.
If you have any issues, please contact with haitao.lin@nlpr.ia.ac.cn