Are GANs overkill for NLP?
Authors: David Alvarez-Melis, Vikas Garg, Adam Kalai
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Some preliminary empirical evidence is also provided to substantiate our theoretical analyses.We provide empirical validation of Lemma 3. |
| Researcher Affiliation | Collaboration | David Alvarez-Melis Microsoft Research daalvare@microsoft.com Vikas Garg Yai Yai Ltd and Aalto University vgarg@csail.mit.edu Adam Tauman Kalai Microsoft Research adam@kal.ai |
| Pseudocode | Yes | Algorithm 1: Boosted weak distinguishers. |
| Open Source Code | No | The paper states: 'We train a pair of text generator and discriminator using a publicly available implementation4 of Seq GAN [63].' with footnote 4 linking to 'https://github.com/Lantao Yu/Seq GAN'. This is a third-party implementation that the authors used, not their own source code for the methodology presented in the paper. |
| Open Datasets | No | The paper mentions 'English text corpus' and refers to 'language models' generally, but does not provide specific access information (link, DOI, citation with author/year) for the dataset used in its empirical validation. |
| Dataset Splits | No | The paper does not provide specific training/validation/test dataset splits, only mentioning 'default Seq GAN settings' for network capacities and language configuration. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU models, CPU models, memory) used for running its experiments. It only generally refers to 'computational resources'. |
| Software Dependencies | No | The paper mentions using 'Seq GAN' but does not provide specific version numbers for this or any other key software dependencies (e.g., programming languages, libraries, frameworks) required for replication. |
| Experiment Setup | No | The paper states: 'The results... correspond to the default Seq GAN settings in terms of network capacities and language configuration (maximum sequence length=20, vocabulary size=5000)'. While this gives some general configuration, it does not provide concrete hyperparameter values or detailed system-level training settings specific to their experiments. |