[HN Gopher] Towards A Token-Free Future In NLP (2022)
___________________________________________________________________
Towards A Token-Free Future In NLP (2022)
Author : optimalsolver
Score : 5 points
Date : 2023-01-30 19:14 UTC (3 hours ago)
(HTM) web link (peltarion.com)
(TXT) w3m dump (peltarion.com)
| f_devd wrote:
| I've actually been working on a Unicode only code prediction LM,
| and it already works pretty well the main issues with Unicode
| models as also found in the article is the large sequence length
| required compared a sentencePiece or BPE tokenizer. The current
| direction to resolve this seems to be to use structured state
| spaces (S4/DSS) models which scale linearly along sequence length
| compared to O(n^2) for transformers. I haven't read the
| charformer paper yet but if it does a dynamic pooling of the
| tokens it could be a promising step to have this same
| functionality in transformer models.
| deepsquirrelnet wrote:
| I work in an industry where most of the text is not natural --
| super wide vocabulary, lots of symbols and abbreviations. I built
| some utf-8 tokenized models that do very well in comparison to
| sentence/word piece tokenizers.
|
| Unless your corpus fits the tokenizer vocabulary very well, I
| think there are many cases where "token free" models can be
| advantageous. You only have to determine how to deal with the
| expansion of the input dimension and how it might blow up your
| attention costs.
___________________________________________________________________
(page generated 2023-01-30 23:02 UTC)