TableLingo

A Data-to-Text Generation Library built on top of Huggingface Transformers and Datasets

Docs

• Paper with findings
• Experiment Protocol
• Literature Review

Setup

pip install -r requirements.txt

Datasets currently supported

• DART
• ToTTo

Models currently supported

• BART
• T5
• Roberta2Roberta shared
• Bert2Bert shared
• Bert2GPT

Linearization strategies

0. <TRIPLE>...</TRIPLE> <TRIPLE>...</TRIPLE> ..... <TRIPLE>...</TRIPLE>
1. permutations of strategy 0 with permutations controlled by max_permutations
2. hierarchical linearization of triple
3. hierarchical linearization of triple with permutations

Training and Evaluation

bash train.sh

or

python pipeline/main.py \
--data_dir ../data/dart/data/v1.1.1/ \
--dataset DART \
--linearize True \
--linearize_strategy 0 \
--run_id test \
--model BART \
--batch_size 16 \
--use_mixed_precision True \
--num_epochs 1 \
--overwrite True