Growing Together: Modeling Human Language Learning With n-Best Multi-Checkpoint Machine Translation

TL;DR

This paper presents a novel approach to machine translation by ensemble modeling across multiple training checkpoints to mimic human language learning stages, improving translation quality.

Contribution

It introduces a multi-checkpoint ensemble method that leverages different training stages to enhance translation fluency and accuracy in language learning models.

Findings

Achieved 37.57 macro F1 score on English-Portuguese translation

Outperformed baseline Amazon translation system with 21.30 macro F1

Demonstrated the effectiveness of multi-checkpoint ensemble approach

Abstract

We describe our submission to the 2020 Duolingo Shared Task on Simultaneous Translation And Paraphrase for Language Education (STAPLE) (Mayhew et al., 2020). We view MT models at various training stages (i.e., checkpoints) as human learners at different levels. Hence, we employ an ensemble of multi-checkpoints from the same model to generate translation sequences with various levels of fluency. From each checkpoint, for our best model, we sample n-Best sequences (n=10) with a beam width =100. We achieve 37.57 macro F1 with a 6 checkpoint model ensemble on the official English to Portuguese shared task test data, outperforming a baseline Amazon translation system of 21.30 macro F1 and ultimately demonstrating the utility of our intuitive method.