hyeontae / context-free-grammar

Evaluating the Ability of LSTMs to Learn Context-Free Grammars

We are trying to reproduce the results of the paper 'Evaluating the Ability of LSTMs to Learn Context-Free Grammars' presented at ACL 2018.

Future plan

1. Construct a corpus with the same statistics with that used in the paper.
2. Train a vanilla sequence-to-sequence model to reproduce the behaviour observed in the paper.
3. Extend the result by trying out different models for machine translation.

To-do

• Modify the distance function
• Compare performances of various sequence-to-sequence vanilla seq2seq models
• Compare performances of various sequence-to-sequence attentional models
• Generate random strings with non-bracket characters such as a,b, and c.
• Generate random strings with unmatching brackets (e.g. '[[}]', '{]')

Memo

1. Please check the "Plus_Model/train_units_to_error_rate_and_accuracy.ipynb" file. (written by Hyeontae - 19/01/15)
2. Unmatching data is not optimized. (written by Hyeontae - 19/02/28)