Back to Blog
Monolingual for 10.6.86/17/2023 ![]() If a benchmark already exists for a dataset/task pair you enter, you’ll see a link appear.Note that you can use parentheses to highlight details, for example: BERT Large (12 layers), FoveaBox (ResNeXt-101), EfficientNet-B7 (NoisyStudent). What are the model naming conventions? Model name should be straightforward, as presented in the paper. ImageNet on Image Classification already exists with metrics Top 1 Accuracy and Top 5 Accuracy. You should check if a benchmark already exists to prevent duplication if it doesn’t exist you can create a new dataset. Then choose a task, dataset and metric name from the Papers With Code taxonomy. You can manually edit the incorrect or missing fields. How do I add a new result from a table? Click on a cell in a table on the left hand side where the result comes from. Help! Don’t worry! If you make mistakes we can revert them: everything is versioned! So just tell us on the Slack channel if you’ve accidentally deleted something (and so on) - it’s not a problem at all, so just go for it! I’m editing for the first time and scared of making mistakes. Where do referenced results come from? If we find referenced results in a table to other papers, we show a parsed reference box that editors can use to annotate to get these extra results from other papers. Where do suggested results come from? We have a machine learning model running in the background that makes suggestions on papers. Blue is a referenced result that originates from a different paper. What do the colors mean? Green means the result is approved and shown on the website. A result consists of a metric value, model name, dataset name and task name. What are the colored boxes on the right hand side? These show results extracted from the paper and linked to tables on the left hand side. It shows extracted results on the right hand side that match the taxonomy on Papers With Code. What is this page? This page shows tables extracted from arXiv papers on the left-hand side. Experiments show that adapter tuning offer competitive results to full fine-tuning, while being much more parameter-efficient. Starting from different pre-trained models (a multilingual ST trained on parallel data or a multilingual BART (mBART) trained on non-parallel multilingual data), we show that adapters can be used to: (a) efficiently specialize ST to specific language pairs with a low extra cost in terms of parameters, and (b) transfer from an automatic speech recognition (ASR) task and an mBART pre-trained model to a multilingual ST task. While adapter tuning was investigated for multilingual neural machine translation, this paper proposes a comprehensive analysis of adapters for multilingual speech translation (ST). ![]() Adapter tuning consists in freezing pretrained parameters of a model and injecting lightweight modules between layers, resulting in the addition of only a small number of task-specific trainable parameters. Lightweight Adapter Tuning for Multilingual Speech TranslationĪdapter modules were recently introduced as an efficient alternative to fine-tuning in NLP.
0 Comments
Read More
Leave a Reply. |