Patching Leaks in the Charformer for Efficient Character-Level Generation
This work addresses efficiency and robustness for character-level models in morphologically rich languages, but it is incremental as it fixes a specific leak in an existing method.
The paper tackled the problem of information leakage in the Charformer's character grouping method when applied to a Transformer decoder, solving this issue to enable character grouping in the decoder. The result showed that Charformer downsampling offers no translation quality benefits over previous methods in NMT but can be trained about 30% faster, with promising performance on English-Turkish translation.
Character-based representations have important advantages over subword-based ones for morphologically rich languages. They come with increased robustness to noisy input and do not need a separate tokenization step. However, they also have a crucial disadvantage: they notably increase the length of text sequences. The GBST method from Charformer groups (aka downsamples) characters to solve this, but allows information to leak when applied to a Transformer decoder. We solve this information leak issue, thereby enabling character grouping in the decoder. We show that Charformer downsampling has no apparent benefits in NMT over previous downsampling methods in terms of translation quality, however it can be trained roughly 30% faster. Promising performance on English--Turkish translation indicate the potential of character-level models for morphologically-rich languages.