Bandits Don't Follow Rules: Balancing Multi-Facet Machine Translation with Multi-Armed Bandits

TL;DR

This paper introduces a multi-armed bandit approach to dynamically balance multi-faceted training data in machine translation, reducing manual tuning and improving system performance across various multi-facet scenarios.

Contribution

It proposes a novel bandit-based method to automatically optimize data facet selection during MT training, enhancing adaptability and reducing manual intervention.

Findings

Bandit learning achieves competitive MT performance.

The approach adapts effectively across multiple facets.

Insights into data selection strategies are provided.

Abstract

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.