Jam or Cream First? Modeling Ambiguity in Neural Machine Translation with SCONES

TL;DR

This paper introduces SCONES, a multi-label classification approach for neural machine translation that models translation ambiguity more effectively than softmax, leading to improved BLEU scores and faster inference.

Contribution

The paper proposes SCONES, a novel loss function and output layer that better handle translation ambiguity, improving translation quality and inference speed in neural machine translation.

Findings

SCONES improves BLEU scores across six translation directions.

Using smaller beam sizes with SCONES speeds up inference by 3.9x.

SCONES mitigates the beam search curse and better models translation ambiguity.

Abstract

The softmax layer in neural machine translation is designed to model the distribution over mutually exclusive tokens. Machine translation, however, is intrinsically uncertain: the same source sentence can have multiple semantically equivalent translations. Therefore, we propose to replace the softmax activation with a multi-label classification layer that can model ambiguity more effectively. We call our loss function Single-label Contrastive Objective for Non-Exclusive Sequences (SCONES). We show that the multi-label output layer can still be trained on single reference training data using the SCONES loss function. SCONES yields consistent BLEU score gains across six translation directions, particularly for medium-resource language pairs and small beam sizes. By using smaller beam sizes we can speed up inference by a factor of 3.9x and still match or improve the BLEU score obtained using softmax. Furthermore, we demonstrate that SCONES can be used to train NMT models that assign the highest probability to adequate translations, thus mitigating the "beam search curse". Additional experiments on synthetic language pairs with varying levels of uncertainty suggest that the improvements from SCONES can be attributed to better handling of ambiguity.