Training customized words

[niranjan8129]

How to train customize word like below

wrong word --------------------- Correct word

G%^oog>.le -------------- Google
$#US12!A ---------------- USA

[mdcramer]

I see you left a few notes, but I'm not sure exactly what you're asking. To call the saved model to get the output, go to the section labeled "Prediction". If you'd like to train custom words you'll need to put that in the training data. I describe how I inject entropy into the Google data file, but you could theoretically inject whatever you like. I hope that helps.

[niranjan8129]

Thank you Sir. I created 2 files in data folder ( sources_letters.txt and target_letters.txt) and then trained model using parameter "small". I am able to creating a model but problem is ,I am not getting expected result and also not able to train word combing with special characters properly . could you please help me on this ?

example below:

sources_letters.txt ---------------------- target_letters.txt

---

$IMXHRÏO&R. ------------------------ INTERIOR
r«10.vl --------------------------- FROM
"Chai? ---------------------------- "Charley,

[mdcramer]

That's a difficult question to answer without digging into your code and the text files you used. Perhaps you don't have enough training data or perhaps you didn't train for long enough. Does your output change at all and does it get closer to the target? If you train for additional epochs does that help? Is your error going down during training?

[niranjan8129]

Sir, I just trained sources_letters.txtand target_letters.txt data . It is working only spelling/word mistake but it is not working if a word has special characters like above example. Also , I marked if I change the test data order after creating the model , it is not working. Do you have few time to look into if you don't feel bad .

[niranjan8129]

correct_words.txt
wrong_words.txt

attached wrong and right word

[mdcramer]

Without digging into this more extensively, here is my best guess: the 'distance' between your correct and incorrect spellings is too great for the neural net to learn a generalized response.

In the 'small' example from my code, the NN is learning how to alphabetize letters, which essentially means learning the alphabet. In the 'large' example I've added a small amount of entropy to each sentence, but the incorrect example isn't that far from the correct example. Here, in essence, the NN has to learn the dictionary plus some other logic, or at the very least has to learn common letter sequences.

With your files the NN is essentially going to have to 'memorize' the translations because there doesn't seem to be a logical or consistent way to move from the incorrect to the correct spellings. (E.g., how would anyone know that '"Chai?' is '"Charlie,'?

I hope that helps.