README.md

---
license: cc-by-sa-4.0
datasets:
- pie/sciarg
- DFKI-SLT/sciarg
language:
- en
---

This is an argument structure prediction model for the scientific domain. It is a pointer network based on 
[A Generative Model for End-to-End Argument Mining with Reconstructed Positional Encoding and Constrained Pointer Mechanism 
(Bao et al., EMNLP 2022)](https://aclanthology.org/2022.emnlp-main.713/). Given a plain input text, the model generates 
in one go tuples that represent argumentative relations, e.g. of type `supports` or `attacks`, between a pair of 
Argumentative Discourse Units (ADUs). Each ADU is defined by start- and end-offsets and a is also typed (`background_claim`, 
`own_claim`, or `data`). 

However, this is a full reimplementation of the model
within the [PyTorch-IE](https://github.com/ArneBinder/pytorch-ie) framework. The model source code can be 
found in the [pie-modules](https://github.com/ArneBinder/pie-modules) repository. The model was trained with the 
[PyTorch-IE-Hydra-Template](https://github.com/ArneBinder/pytorch-ie-hydra-template-1) on the 
[SciArg dataset](https://aclanthology.org/W18-5206/), see [here](https://huggingface.co/datasets/pie/sciarg) for 
further information and an integration into [pie-datasets](https://github.com/ArneBinder/pie-datasets). Further 
information regarding the training setup and model performance can be found in the [config.yaml](config.yaml), 
in the [wandb-metadata.json](wandb-metadata.json), and in [wandb-summary.json](wandb-summary.json). 
([link to private W&B run](https://wandb.ai/arne/dataset-sciarg-task-ner_re-v0.3-training/runs/wr3bg4la))

You can try out the model in [this HF space](https://huggingface.co/spaces/ArneBinder/sam-pointer-bart-base-v0.3).