MeetingBank Summarization with BART

Fine-tuning facebook/bart-large-xsum on the MeetingBank dataset to generate concise meeting summaries. This repo contains a simple, script-based pipeline from preprocessing to training and evaluation.

Repository Structure

• 01_data_preprocessing.py: Loads MeetingBank, tokenizes transcripts and summaries, and saves a tokenized dataset.
• 02_fine_tuning_BART.py: Fine-tunes BART on the tokenized dataset and saves checkpoints and logs.
• 03_evaluation_and_comparison.py: Compares the fine-tuned model with the base model using ROUGE and exports side-by-side outputs.
• Slides/: Presentation screenshots for context and visualization.
• results/: Saved experiment outputs (loss logs and evaluation exports).

Setup

• Python 3.9+ recommended
• GPU strongly recommended for training and evaluation

Install dependencies:

pip install -r requirements.txt

Workflow

1. Preprocess and tokenize the dataset:

python 01_data_preprocessing.py

2. Fine-tune BART (default configuration uses training_args_v2_2):

python 02_fine_tuning_BART.py

3. Evaluate and compare against the base model:

python 03_evaluation_and_comparison.py

Outputs

• Tokenized dataset saved to data/processed/tokenized_meetingbank
• Models saved to models/ (path configured in 02_fine_tuning_BART.py)
• Training logs and CSV exports in results/

Notes

• Make sure fine_tuned_model_path in 03_evaluation_and_comparison.py matches the model output path used in training.
• The evaluation script runs on the full test split by default and can take a while on CPU.
• If you want smaller, faster experiments, use the provided half_train_data, tenth_train_data, or hundredth_train_data in 02_fine_tuning_BART.py.

Dataset

• MeetingBank (Hugging Face Hub): huuuyeah/meetingbank

Citations

1. MeetingBank: A Benchmark Dataset for Meeting Summarization
   Yebowen Hu, Tim Ganter, Hanieh Deilamsalehy, Franck Dernoncourt, Hassan Foroosh, Fei Liu
   In main conference of Association for Computational Linguistics (ACL'23), Toronto, Canada.

2. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
   Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, Luke Zettlemoyer

License

• MIT (see LICENSE).