Smoothing and Shrinking the Sparse Seq2Seq Search Space

TL;DR

This paper introduces entmax-based sparse sequence models that effectively address the length bias in seq2seq models, improving translation and linguistic task performance by shrinking the search space and generalizing label smoothing.

Contribution

It demonstrates the effectiveness of entmax models in solving the length bias problem and generalizes label smoothing to Fenchel-Young losses, achieving state-of-the-art results.

Findings

Entmax models remove the 'cat got your tongue' problem in translation.

Label smoothing is generalized to Fenchel-Young losses.

Models achieve state-of-the-art on multilingual grapheme-to-phoneme conversion.

Abstract

Current sequence-to-sequence models are trained to minimize cross-entropy and use softmax to compute the locally normalized probabilities over target sequences. While this setup has led to strong results in a variety of tasks, one unsatisfying aspect is its length bias: models give high scores to short, inadequate hypotheses and often make the empty string the argmax -- the so-called cat got your tongue problem. Recently proposed entmax-based sparse sequence-to-sequence models present a possible solution, since they can shrink the search space by assigning zero probability to bad hypotheses, but their ability to handle word-level tasks with transformers has never been tested. In this work, we show that entmax-based models effectively solve the cat got your tongue problem, removing a major source of model error for neural machine translation. In addition, we generalize label smoothing, a critical regularization technique, to the broader family of Fenchel-Young losses, which includes both cross-entropy and the entmax losses. Our resulting label-smoothed entmax loss models set a new state of the art on multilingual grapheme-to-phoneme conversion and deliver improvements and better calibration properties on cross-lingual morphological inflection and machine translation for 6 language pairs.