Just a query, have you tried triplet loss or lossless triplet loss as I think that would produce better embeddings as we are providing fewer examples and the clusters formed will be visually better.
Looking for this feature if it is the right approach.
Also, this is will only handle non-word error and won't work for real word error if error-correction is your aim.
Although, I liked the approach and guess it is similar to fastText from Facebook, here you are using RNN whereas they use n-grams. As I am a beginner considering all this, can you guide me with which method is preferable and reason?
fjsj
commented
3 years ago
Just a query, have you tried triplet loss or lossless triplet loss as I think that would produce better embeddings as we are providing fewer examples and the clusters formed will be visually better. Looking for this feature if it is the right approach.
Also, this is will only handle non-word error and won't work for real word error if error-correction is your aim.
Although, I liked the approach and guess it is similar to fastText from Facebook, here you are using RNN whereas they use n-grams. As I am a beginner considering all this, can you guide me with which method is preferable and reason?