Deep Learning for Abstract Argumentation Semantics
Authors: Dennis Craandijk, Floris Bex
IJCAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experimental results demonstrate that the AGNN can almost perfectly predict the acceptability under different semantics and scales well for larger argumentation frameworks. |
| Researcher Affiliation | Collaboration | 1 National Police Lab AI, Netherlands Police 2 Information and Computing Sciences, Utrecht University 3 Institute for Law, Technology and Society, Tilburg University {d.f.w.craandijk, f.j.bex}@uu.nl |
| Pseudocode | No | The paper describes the AGNN model's operation and message passing steps in text, but it does not include a dedicated pseudocode block or algorithm section. |
| Open Source Code | Yes | We publish our code at https://github.com/Dennis Craandijk/ DL-Abstract-Argumentation. |
| Open Datasets | No | We generate a variety of challenging argumentation frameworks by sampling from the following AF generators from the International Competition on Computational Models of Argumentation [Gaggl et al., 2020]: AFBench Gen2, AFGen Benchmark Generator, Grounded Generator, Scc Generator, Stable Generator. |
| Dataset Splits | Yes | We generate a test and validation dataset of size 1000 with AFs containing |A| = 25 arguments, and a training dataset of a million AFs where the number of arguments per AF is sampled randomly between 5 |A| 25 (to accelerate the learning). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running the experiments. |
| Software Dependencies | No | Ground-truth labels are determined based on extensions obtained with the sound and complete µ-toksia solver [Niskanen and J arvisalo, 2019]. A specific version for µ-toksia is not provided, nor for AdamW, LSTM, or the overall framework (e.g., PyTorch version). |
| Experiment Setup | Yes | The dimensions of the embedding and all hidden neural layers are d = 128. The model is run for T = 32 message passing steps. We train our model in batches containing 50 graphs (approximately 750 nodes) using the Adam W optimiser [Loshchilov and Hutter, 2019] with a cosine cyclical learning rate [Smith, 2017] between 2e 4 and 1e 7, ℓ2 regularisation of 1e 9 and clip the gradients by global norm with a 0.5 clipping ratio [Pascanu et al., 2013]. |