Hello, I have tried using a basic seq2seq LSTM architecture and trained it for spell-checking purposes but performance is not satisfactory. I am not looking to make the sentences themselves grammatically correct, only spell-checking the words (which might sometimes need context of the sentence to understand what word was meant to be typed). what are state of the art models that can be trained/fine-tuned for this purpose?
#Spell-checking
inland dragon
devout gyro
Why not look at prefix trees for auto correct?
Alternatively, just keep a literal dictionary and compute something like levenstien distance between words not found in the dictionary and the “closest” matching term?
Then just have a model like BERT trained to take that list and output a label for which word is “most probable” given the context of the sentence.