Skip to content

wmt-2014.md

Latest commit

 

History

History
31 lines (26 loc) · 1.37 KB

wmt-2014.md

File metadata and controls

31 lines (26 loc) · 1.37 KB
Raw
sources
summaries/attention-is-all-you-need.md
type Event
description 2014 machine translation benchmark and data release.

WMT 2014

Overview

WMT 2014 is the machine translation benchmark year used in [[summaries/attention-is-all-you-need]] to evaluate the Transformer on English-German and English-French translation. In the paper, it serves as the standard comparison point for prior state-of-the-art systems and for reporting BLEU scores.

Key facts from the paper

  • The paper evaluates on the WMT 2014 English-German task and the WMT 2014 English-French task.
  • These benchmarks are used to compare the Transformer against prior sequence-to-sequence systems.
  • The Transformer achieves 28.4 BLEU on English-German and 41.8 BLEU on English-French.
  • The paper reports that these results come at a fraction of the training cost of earlier models.

Related pages

  • [[entities/wmt-2014-english-german]]
  • [[entities/wmt-2014-english-french]]
  • [[concepts/transformer-models]]
  • [[concepts/attention-mechanisms]]
  • [[entities/tensor2tensor]]
  • [[entities/google]]
  • [[entities/google-brain]]
  • [[entities/google-research]]
  • [[entities/nips-2017]]
  • [[summaries/attention-is-all-you-need]]

Notes

In this document, WMT 2014 refers to the benchmark settings and datasets used for machine translation evaluation rather than a single isolated conference event.