Post-processing to improve the predicted DSL

[emmcohen]

Dear Emil,

Thanks for this great job, I really appreciate the project.

Concerning the bootstrap version, I was wondering if you used a post-processing step to improve the DSL prediction.
In fact, when obtaining a DSL predicted by the model, then, the compiler used for HTML generation is really sensitive to the quality of the prediction.

For instance, the following cases will lead to a 'None' HTML or 'parsing error' in the compiler :

• the prediction contains 2 successives ' , '
• 'text text' without ' , ' to separate them
• succession of ' , ' and ' } '

In general, how do you deal with errors in DSL prediction that could be critical ?

Thanks a lot,