4 ms·
Aren't most autocorrect transformer-based?
by stymaar 2mo ago
Aren't most autocorrect transformer-based?
- sebastiennight 2mo agoHistorically, no. The simplest way to build autocorrect is to have a dictionary, look it up for each word, and suggest the "closest" dictionary word if a typed word is absent from the dictionary. It has existed way before transformers and on platforms that couldn't dream of running even the smallest transformer.
- sublinear 2mo agoWhat do you mean by "look it up" if not measure the closest dictionary word by edit distance? I think you'll find the broader idea of word vectorization to be remarkably similar to what text transformers do. https://en.wikipedia.org/wiki/Edit_distance https://en.wikipedia.org/wiki/Edit_distance
- sebastiennight 2mo agoYes, I meant, find the closest word by edit distance. I'm unsure where vectorization would factor in here?
- sublinear 2mo agoWhen I made this comment, I was thinking of edit distance embedding (like everyone else in this reply chain) while you're thinking of BK-trees. We're just talking past each other. I agree that a tree is the historical solution. I didn't catch that part.