contextualSpellCheck: Misbehavior?

[Halvani]

import spacy
import contextualSpellCheck

nlp = spacy.load('en_core_web_sm')
contextualSpellCheck.add_to_pipe(nlp)

doc = nlp('Conference on Empirical Methodsin Natural Language Processing')
print(doc._.outcome_spellCheck)

# Conference on Empirical Analysis Natural Language Processing

I'm not sure if a spell checker should actually perform a paraphrase.
Shouldn't it just correct the spelling errors? ...so that the correct output becomes:

"Conference on Empirical Methods in Natural Language Processing"

[ljvmiranda921]

Hi @Halvani , this might be due to the fact that contextualSpellChecker uses BERT embeddings (i.e., checking the context of a given token) to estimate how a word should be corrected. You may have a shot by preprocessing the text a bit to better guide the spell-checker, trying out a rules-based approach (perhaps by using other tools like autocorrect), or experimenting with some of the spell checker's parameters. For the last suggestion, perhaps you can try playing around the edit distance or the BERT model? Just to see how strict it will be when correcting your text.