This model translates text from Russian to English. It expects UTF-8 Russian text as input. It outputs English text. The model was developed by Facebook.
Create a Modzy account to get started →
40% Bleu Score
This model is trained on several corpora that are a part of the WMT-19. These consist of the parallel corpora that are in the Paracrawl, Common Crawl, news-commentary, Yandex, Wiki-titles and the UN datasets. This model achieves a Bleu score of 40.0 which is state of the art for 2019. The Bleu is a commonly used metric for translation tasks and it is a way of comparing model generated text to a gold standard to see how similar they are.
The Bleu is a method for assessing the quality of text that has been machine-translated from one language to another. The closer the machine translation is to expert human translation, the better the score.
This model uses the Facebook Fairseq sequence modeling toolkit. This model is based on the big Transformer architecture as implemented in Fairseq.
In a standard Transformer model, there are stacked encoders which communicate with stacked decoders. As a part of these components, there are sublayers which implement Attention. This attempts to enable the model to look at other parts of the input and output sentence as it translates instead of just looking at the individual word. The input text is converted to embeddings which are then fed into the layers of the model.
The authors found that by increasing the feedforward network size they were able to achieve a reasonable improvement in performance while maintaining a manageable network size.
This model is trained on several corpora that are a part of the WMT-19. These consist of the parallel corpora that are in the Paracrawl, Common Crawl, news-commentary, Yandex, Wiki-titles and the UN datasets. The data had several preprocessing steps including language identification, large scale back-translation, ensembling, re-ranking as well as training first on lower quality datasets and then fine-tuning on higher quality ones.
This model was validated on the test set of the WMT-19 dataset and achieves a Bleu score of 40.0. This test set was created out of headlines from September-November 2018. Additionally, in the WMT-19 competition there was also human evaluation.
The input(s) to this model must adhere to the following specifications:
This model will output the following:
Get a video demo of Modzy, the ModelOps and MLOps software platform that businesses use to deploy, integrate, run, and monitor AI—anywhere.