vinhkhuc/MemN2N-babi-python

Question about data_util.py

Closed this issue · 3 comments

HanyuGuo commented

I find that data_util.py ignores comma while parsing babi task. That is, if the answer of a question is s,w(denotes south and west in positional_reasoning), this parser would treat them as a unit. Furthermore, while asking what stuff someone has and the answer would be "milk,pen,hat" or something like that, the parser would treat them as a unit as well. Is that a correct way?

vinhkhuc commented

Hi guohanyu123,
Sorry for the late reply. Since MemN2N assumes that each answer is just one word, "s,w" and "milk,pen,hat" should be treated as units.

HanyuGuo commented

Thanks for your reply, vinhkhuc. So apart from MemN2N, can I say that adding an RNN decoder for seq2seq output would provide a better performance? I am just wondering.

vinhkhuc commented

That's an interesting question. I'm not sure if anyone has tried it before. Also, using grammar tree parser + word embedding may help.
Besides that, adding gradient noise can improve the accuracy of memory networks: http://arxiv.org/abs/1511.06807.