AlignSum: Data Pyramid Hierarchical Fine-tuning for Aligning with Human Summarization Preference

Data Preparation

You need to download the original XSum and CNNDM, put them in data/original

element-aware-cnndm and element-aware-xsum can be download from https://github.com/Alsace08/SumCoT, and put them in data/cnn_dailymail and data/xsum.

Element-Aware Dataset Sample and Data Pyramid Construction

# Select 100 samples as human-annotated data and another 100 samples as test data.
# You can also use the sampled data from our experiments located in the sampled-data directory.
python dp/sample_user_specific_data.py --dataset cnn_dailymail

python dp/generate_GSG.py --dataset cnn_dailymail
CUDA_VISIBLE_DEVICES=0 python dp/generate_LLM.py --dataset cnn_dailymail
python dp/merge_and_gaussian.py --dataset cnn_dailymail

Two-stage Hierarchical Fine-tuning

bash scripts/cnn_ft_stage1.sh
bash scripts/cnn_ft_stage2.sh

Automatic Evaluation

python bs_score.py

You may need to change the result path.