Bandits Don't Follow Rules: Balancing Multi-Facet Machine Translation with Multi-Armed Bandits
This addresses the challenge for MT developers in manually designing data schedules, though it is incremental as it applies an existing bandit method to a known data balancing issue.
The paper tackled the problem of balancing multi-faceted training data in machine translation, such as domains or quality levels, by using a multi-armed bandit to dynamically select facets, resulting in competitive MT systems across tasks.
Training data for machine translation (MT) is often sourced from a multitude of large corpora that are multi-faceted in nature, e.g. containing contents from multiple domains or different levels of quality or complexity. Naturally, these facets do not occur with equal frequency, nor are they equally important for the test scenario at hand. In this work, we propose to optimize this balance jointly with MT model parameters to relieve system developers from manual schedule design. A multi-armed bandit is trained to dynamically choose between facets in a way that is most beneficial for the MT system. We evaluate it on three different multi-facet applications: balancing translationese and natural training data, or data from multiple domains or multiple language pairs. We find that bandit learning leads to competitive MT systems across tasks, and our analysis provides insights into its learned strategies and the underlying data sets.