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). 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 framework. The model source code can be found in the pie-modules repository. The model was trained with the PyTorch-IE-Hydra-Template on the SciArg dataset, see here for further information and an integration into pie-datasets. Further information regarding the training setup and model performance can be found in the config.yaml, in the wandb-metadata.json, and in wandb-summary.json. (link to private W&B run)

You can try out the model in this HF space.

Downloads last month
51
Inference API
Unable to determine this model’s pipeline type. Check the docs .

Datasets used to train ArneBinder/sam-pointer-bart-base-v0.3