Summary of On the Shortcut Learning in Multilingual Neural Machine Translation, by Wenxuan Wang et al.
On the Shortcut Learning in Multilingual Neural Machine Translation
by Wenxuan Wang, Wenxiang Jiao, Jen-tse Huang, Zhaopeng Tu, Michael R. Lyu
First submitted to arxiv on: 15 Nov 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The study investigates the off-target issue in multilingual neural machine translation (MNMT), where a model mistakenly translates one language into another. The authors attribute this issue to overfitting of shortcuts, which biases the model to translate non-centric languages into centric languages instead. They analyze learning dynamics and find that shortcut learning occurs in the later stages of training, accelerated by multilingual pretraining. To eliminate these shortcuts, they propose a simple training strategy that removes instances inducing shortcut learning during the latter stage of training. This approach improves zero-shot translation performance without adding data or computational costs. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This study looks at how machine translators can get confused when trying to translate between different languages. The researchers found that this happens because the translator is overfitting, which means it’s getting too good at a specific task and forgetting about other tasks. They think this problem is caused by the way the translator learns shortcuts, or quick ways to do things. To fix this, they came up with a simple new way of training the translator that helps it remember what it learned earlier on. |
Keywords
» Artificial intelligence » Overfitting » Pretraining » Translation » Zero shot
