LLM-T2T

The data and code for the EMNLP 2023 industry-track paper Investigating Table-to-Text Generation Capabilities of LLMs in Real-World Information Seeking Scenarios. This research investigates the table-to-text capabilities of different LLMs using four datasets within two real-world information seeking scenarios. It demonstrates that high-performing LLMs, such as GPT-4, can effectively serve as table-to-text generators, evaluators, and feedback generators.

Environment Setup

The code is tested on the following environment:

python 3.9.18
CUDA 12.1
run pip install -r requirements.txt to install all the required packages

Data and Model Output

The dataset we used can be found in the data folder, and the model outputs are in the output folder.

Table-to-Text Generation by GPT-series Models:

At first, modify the bash script GPT_T2T_generation.sh. The description of all the arguments used in the script can be found in GPT_T2T_generation.py.

Commands to generate text directly by the LLM (RQ1):

export CUDA_VISIBLE_DEVICES=0,1; 
export PYTHONPATH=`pwd`;
python src/GPT_T2T_generation.py \
    --api_org  \
    --api_key  \
    --engine gpt-3.5-turbo \
    --temperature 0.7 \
    --top_p 1.0 \
    --mode direct \
    --direct_mode two_shot_CoT \
    --dataset FeTaQA \
    --output_path output/FeTaQA/ \
    --num_limit 500 # number of tables used for generation

Commands to improve the output of the finetuned small models by the LLM (RQ3):

export CUDA_VISIBLE_DEVICES=0,1; 
export PYTHONPATH=`pwd`;
python src/GPT_T2T_generation.py \
    --api_org  \
    --api_key  \
    --engine gpt-3.5-turbo \
    --temperature 0.7 \
    --top_p 1.0 \
    --mode improve \
    --dataset LogicNLG \
    --finetuned_model_path output/LogicNLG/original/100tables/GPT2_100tables.json \
    --num_paths 3 \
    --output_path output/LogicNLG/ \
    --num_limit 500 # number of tables used for generation

Table-to-Text Generation by Open-Source Models:

At first, modify the bash script open_src_model_T2T_generation.sh. The description of all the arguments used in the script can be found in open_src_model_T2T_generation.py.

Commands to generate text directly by the LLM (RQ1):

export CUDA_VISIBLE_DEVICES=0,1; 
export PYTHONPATH=`pwd`;
python src/open_src_model_T2T_generation.py \
    --model_type llama-7b \
    --model_name_or_path huggyllama/llama-7b \
    --mode direct \
    --direct CoT \
    --dataset FeTaQA \
    --output_path output/FeTaQA \
    --num_limit 500 # number of tables used for generation

Commands to improve the output of the finetuned small models by the LLM (RQ3):

export CUDA_VISIBLE_DEVICES=0,1; 
export PYTHONPATH=`pwd`;
python src/open_src_model_T2T_generation.py \
    --model_type llama-7b \
    --model_name_or_path huggyllama/llama-7b \
    --mode improve \
    --dataset LogicNLG \
    --finetuned_model_path data/LogicNLG/original/100tables/GPT2_100tables.json \
    --output_path output/LogicNLG \
    --num_limit 500 # number of tables used for generation

Contact

For any issues or questions, kindly email us at: Yilun Zhao (yilun.zhao@yale.edu), Haowei Zhang (haowei.zhang@tum.de) or Shengyun Si (shengyun.si@tum.de).

Citation

@inproceedings{zhao-etal-2023-investigating,
    title = "Investigating Table-to-Text Generation Capabilities of Large Language Models in Real-World Information Seeking Scenarios",
    author = "Zhao, Yilun  and
      Zhang, Haowei  and
      Si, Shengyun  and
      Nan, Linyong  and
      Tang, Xiangru  and
      Cohan, Arman",
    editor = "Wang, Mingxuan  and
      Zitouni, Imed",
    booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.emnlp-industry.17",
    doi = "10.18653/v1/2023.emnlp-industry.17",
    pages = "160--175",
}