Hi-ArG

Code repository of Hi-ArG: Exploring the Integration of Hierarchical Argumentation Graphs in Language Pretraining.

Requirements

Please use Python 3.10 or above to run all codes. You can install the required python packages with pip install -r requirements.txt.

Configure AMRLib

Additionally, if you want to generate Hi-ArG, you need to follow the extra steps below to configure AMRLib, after installing the requirements:

• Install SpaCy model: python3 -m spacy download en_core_web_sm;
• Install and compile fast_align (https://github.com/clab/fast_align), putting it in your PATH;
• Install parse_xfm_bart_large parse model for AMRLib (https://github.com/bjascob/amrlib-models).

Please refer to https://amrlib.readthedocs.io/en/latest/install/ for more information.

Code Usage

Hi-ArG Generation

To generate Hi-ArG for args.me, ArgKP-2021 or IAM, navigate to argsme, argkp2021 or iam_cesc, run mkdir -p data/raw and put raw data inside data/raw. Next, run ./work.sh, then you can find generated Hi-ArG in data/final.

You may want to add parameters after python 10_parse.py to make use of GPU and adjust other settings when parsing AMR graphs. See the output of python 10_parse.py -h for more details.

Specifications for raw data files:

• For args.me, the data file should be named args-me-1.0-cleaned.json; you can download the data file at https://doi.org/10.5281/zenodo.4139439.

• For ArgKP-2021, the data files should be named {part}_{set}.csv, where {part} can be arguments/key_points/labels and {set} be train/dev/test; you can download the data files at https://github.com/ibm/KPA_2021_shared_task (under the kpm_data and test_data folders).

• For IAM, the data files should be named {set}.txt, where {set} can be train/dev/test. You can download the data files at https://github.com/LiyingCheng95/IAM (under the claims folder).

Model Training

Before training models, make sure that you have generated Hi-ArG following the steps in the previous section. You can check this by looking into data/final to see if the generated Hi-ArG are there.

You also need to convert raw graphs into trainable data. Assume that you are already under argsme, argkp2021 or iam_cesc, navigate to prepare and run ./work.sh to generate data for fine-tuning.

Further Pre-training

You need to get further pre-trained models before fine-tuning them for downstream tasks. To do so, first navigate to argsme/pretrain and run ./prepare.sh. This will create necessary symbolic links and cache files for the training data. Next, run ./work.sh or python train.py with necessary arguments (see the content of work.sh and the output of python train.py -h for more details) to further pre-train models.

Under Linux, you can share cached data across parallel trainings by copying cache files in the cache folder to /dev/shm. You can check the sources of all cache files in cache_info/*.jsonl to decide which files to copy.

Fine-tuning

The fine-tuning procedure is essentially the same as pre-training, except that no cache files are created: navigate to argkp2021/finetune or iam_cesc/finetune, run ./prepare.sh and then ./work.sh or python train.py with necessary arguments. This will also generate evaluation results in evaluate; you can obtain evaluation statistics by running ./summary.sh afterwards and looking into summary/per_pretrain or summary/grand_total.

ChatGPT Experiments

To run ChatGPT experiments, make sure that raw data files are prepared under argkp2021/generate/data/raw or iam_cesc/generate/data/raw, and then run ./work.sh {API_KEY} under gpt/argkp2021 or gpt/iam_cesc; replace {API_KEY} with your own OpenAI API key. You can also specify different models or base URLs by directly running python predict.py with necessary parameters.

Citation

@Article{Liang2023Hi,
  author        = {Liang, Jingcong and Ye, Rong and Han, Meng and Zhang, Qi and Lai, Ruofei and Zhang, Xinyu and Cao, Zhao and Huang, Xuanjing and Wei, Zhongyu},
  title         = {Hi-ArG: Exploring the Integration of Hierarchical Argumentation Graphs in Language Pretraining},
  year          = {2023},
  archiveprefix = {arXiv},
  eprint        = {2312.00874},
  primaryclass  = {cs.CL},
}