SLTUNET: A Simple Unified Model for Sign Language Translation

Authors: Biao Zhang, Mathias Müller, Rico Sennrich

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show in experiments that SLTUNET achieves competitive and even state-of-the-art performance on PHOENIX-2014T and CSL-Daily when augmented with MT data and equipped with a set of optimization techniques. We further use the DGS Corpus for end-to-end SLT for the first time. It covers broader domains with a significantly larger vocabulary, which is more challenging and which we consider to allow for a more realistic assessment of the current state of SLT than the former two. Still, SLTUNET obtains improved results on the DGS Corpus.
Researcher Affiliation Academia 1 School of Informatics, University of Edinburgh 2 Department of Computational Linguistics, University of Zurich
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/bzhang Go/sltunet.
Open Datasets Yes We work on three SLT datasets: PHOENIX-2014T (Camgoz et al., 2018), CSL-Daily (Zhou et al., 2021), and DGS3-T.
Dataset Splits Yes The split contains 60,306, 967, and 1,575 samples in the train, dev, and test set, respectively (see Table 1 and Appendix A.2 for details).
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications) used for running its experiments.
Software Dependencies No The paper mentions software like Moses and Sacre BLEU with citations but does not provide specific version numbers for them or other key software dependencies used in the experiments.
Experiment Setup Yes We experiment with Transformer (Vaswani et al., 2017) and start our analysis with a Baseline system optimized on Sign2Text alone with the following configurations: encoder and decoder layers of N S enc = 2, N P enc = 0 and Ndec = 2 respectively, model dimension of d = 512, feed-forward dimension of dff = 2048, attention head of h = 8, and no CTC regularization. ... We train all SLT models using Adam (β1 = 0.9, β2 = 0.998) (Kingma & Ba, 2015) with Noam learning rate schedule (Vaswani et al., 2017), a label smoothing of 0.1 and warmup step of 4K. We employ Xavier initialization to initialize model parameters with a gain of 1.0.