[HW3] Reasonable accuracy Problem 2.2

[dinhkt]

Hello professor and TAs,
I have questions in problem 2.2 and 2.3

1. In prob 2.2,I used a filtered vocab from Squad, which has 40000 words for training. First I tried to create noise by put a random character to a random position in word, and the accuracy of correction is around 10%. Then I create noise by randomly double a character in the word, and the accuracy is about 25-30%. How should noise be created and what is the expected accuracy?
2. Problem 2.3 is still spelling correction for a word, but now consider the other words in the sentence, right? Can you give a hint how to design that?

[seominjoon]

1. What you tried is reasonable, and other possible ideas are: partial shuffing, missing letter, etc.
2. One way is to just use the same model as 2.2, but you will need a new data for training.

[dinhkt]

Thank you professor,
For question 2, so that we can split the sentence into characters and space and then encode ?

[seominjoon]

Yes, and there is no single right answer.