Repository navigation
fix: Inconsistent Token Alignment Between Fast and Slow Tokenizers - #426
Open
mayuriphad wants to merge 1 commit into
Open
mayuriphad wants to merge 1 commit into
mayuriphad wants to merge 1 commit into
Conversation
…ate alignment with fast tokenizers
Author
|
Hi maintainers, a gentle ping on this PR whenever you have a moment to review. I'm happy to make any changes or add tests if something needs adjusting. Thank you for your time! |
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes explosion/spaCy#13730
When using fast tokenizers (like DeBERTa), the alignment IDs were being duplicated across adjacent spaCy tokens. This happened because char_to_sp_token mapped the trailing whitespace (which isn't part of the oken.text) to the preceding token.
Fast tokenizers often produce offset_mapping that includes the preceding space for wordpieces (e.g. _word). When that space character mapped to the previous token, the wordpiece was aligned to BOTH the previous token and the current token, resulting in duplicate alignments (�lign_lengths = 2).
This patch removes the + 1 from the range loop so that trailing spaces are not mapped to the preceding token. Wordpieces will only align to tokens whose non-whitespace characters they overlap with.