Latent Template Induction with Gumbel-CRFs

Authors: Yao Fu, Chuanqi Tan, Bin Bi, Mosha Chen, Yansong Feng, Alexander Rush

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of our methods with experiments on data-to-text generation and unsupervised paraphrase generation.
Researcher Affiliation Collaboration Yao Fu1, , Chuanqi Tan2, Bin Bi2, Mosha Chen2, Yansong Feng3, Alexander M. Rush4 1ILCC, University of Edinburgh, 2Alibaba Group, 3WICT, Peking Univeristy, 4Cornell University
Pseudocode Yes Algorithm 1 Forward Filtering Backward Sampling, Algorithm 2 Gumbel-CRF (Forward Filtering Backward Sampling with Gumbel-Softmax)
Open Source Code Yes Our code is available at https://github.com/FranxYao/Gumbel-CRF.
Open Datasets Yes For text modeling and data-to-text generation, we use the E2E dataset[50]... For paraphrase generation we follow the same setting as Fu et al. [16], and use the common MSCOCO dataset.
Dataset Splits Yes This dataset contains approximately 42K training, 4.6K validation and 4.6K testing sentences.
Hardware Specification Yes Models tested on Nvidia P100 with batch size 100.
Software Dependencies No The paper mentions various parameters for tuning (e.g., 'h entropy regularization, τ temperature annealing') but does not specify software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x).
Experiment Setup Yes At the beginning of training, to prevent the decoder from ignoring z, we apply word dropout [7]... For optimization, we add a β coefficient to the entropy term, as is in the β-VAE [23]... Models tested on Nvidia P100 with batch size 100.