[ML] Ensure BertTokenizer does not split special tokens
#81254
Merged
Conversation
This file contains 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
This commit changes the way our Bert tokenizer preserves special tokens without splitting them. The previous approach split the tokens first on common punctuation, checked if there were special tokens, and then did further split on all punctuation. The problem with this was that words containing special tokens and non-common punctuation were not handled properly. This commit addresses this by building a trie tree for the special tokens, splitting the input on all punctuation, and then looking up the tokens in the special token trie in order to merge matching tokens back together. Closes elastic#80484
|
Pinging @elastic/ml-core (Team:ML) |
davidkyle
approved these changes
Dec 2, 2021
...ugin/ml/src/main/java/org/elasticsearch/xpack/ml/inference/nlp/tokenizers/TokenTrieNode.java
Outdated
Show resolved
Hide resolved
...gin/ml/src/main/java/org/elasticsearch/xpack/ml/inference/nlp/tokenizers/BasicTokenizer.java
Show resolved
Hide resolved
benwtrent
reviewed
Dec 2, 2021
| @@ -166,12 +166,21 @@ public void testPunctuation() { | |||
|
|
|||
| public void testPunctuationWithMask() { | |||
There was a problem hiding this comment.
Could we verify that neversplit works on something like "this should never split"?
There was a problem hiding this comment.
You mean a never-split token that contains whitespace? If yes, then it doesn't work with that as things are as it's using the same tokenization code as that used for the text input and that splits on whitespace.
benwtrent
approved these changes
Dec 6, 2021
dimitris-athanasiou
added a commit
to dimitris-athanasiou/elasticsearch
that referenced
this pull request
Dec 6, 2021
) This commit changes the way our Bert tokenizer preserves special tokens without splitting them. The previous approach split the tokens first on common punctuation, checked if there were special tokens, and then did further split on all punctuation. The problem with this was that words containing special tokens and non-common punctuation were not handled properly. This commit addresses this by building a trie tree for the special tokens, splitting the input on all punctuation, and then looking up the tokens in the special token trie in order to merge matching tokens back together. Closes elastic#80484
💚 Backport successful
|
elasticsearchmachine
pushed a commit
that referenced
this pull request
Dec 6, 2021
…) (#81371) * [ML] Ensure `BertTokenizer` does not split special tokens (#81254) This commit changes the way our Bert tokenizer preserves special tokens without splitting them. The previous approach split the tokens first on common punctuation, checked if there were special tokens, and then did further split on all punctuation. The problem with this was that words containing special tokens and non-common punctuation were not handled properly. This commit addresses this by building a trie tree for the special tokens, splitting the input on all punctuation, and then looking up the tokens in the special token trie in order to merge matching tokens back together. Closes #80484 * Fix compilation
dimitris-athanasiou
added a commit
to dimitris-athanasiou/elasticsearch
that referenced
this pull request
Dec 7, 2021
…ns (elastic#81254) (elastic#81371)" This reverts commit 9fc8415.
dimitris-athanasiou
added a commit
that referenced
this pull request
Dec 7, 2021
dimitris-athanasiou
added a commit
to dimitris-athanasiou/elasticsearch
that referenced
this pull request
Dec 7, 2021
This fixes a bug introduced by elastic#81254. We are now using a token trie tree to merge tokens belonging to one of the never-split tokens back together. However, if the tokenizer is lower casing, then the merged token will also be lower case and won't be matched against never split tokens that are expected to be in upper case. This commit fixes this by looking up the original text and only merging tokens together when the original text is matching one of the never split tokens.
dimitris-athanasiou
added a commit
that referenced
this pull request
Dec 7, 2021
This fixes a bug introduced by #81254. We are now using a token trie tree to merge tokens belonging to one of the never-split tokens back together. However, if the tokenizer is lower casing, then the merged token will also be lower case and won't be matched against never split tokens that are expected to be in upper case. This commit fixes this by looking up the original text and only merging tokens together when the original text is matching one of the never split tokens.
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.
This commit changes the way our Bert tokenizer preserves
special tokens without splitting them. The previous approach
split the tokens first on common punctuation, checked if there
were special tokens, and then did further split on all punctuation.
The problem with this was that words containing special tokens and
non-common punctuation were not handled properly.
This commit addresses this by building a trie tree for the special
tokens, splitting the input on all punctuation, and then looking up
the tokens in the special token trie in order to merge matching tokens
back together.
Closes #80484