billpei/dmnplustf

Dynamic Memory Network Plus in Tensorflow

This is my implementaion of the super cool paper by Caiming Xiong, Stephen Merity, Richard Socher of MetaMind: DMN for Visual and Textual Question Answering

Many many thanks to @therne and @barronalex for sharing their code - very useful reference.

Gratitude to Richard Socher for sharing knowledge openly via cs224d and Andrej Karpathy (@karpathy) via cs231n!

##Benchmarks Test Error on the Facebook's babi 10k data set for Question Answering:

	Task	This implementation	Xiong et al DMN+
1	Basic factoid QA with single supporting fact	0	0
2	Factoid QA with two supporting facts	0	0.3
3	Factoid QA with three supporting facts	8.04	1.1
4	Two argument relations: subject vs. object	0	0
5	Three argument relations	0.23	0.5
6	Yes/No questions	0	0
7	Counting	4.36	2.4
8	Lists/Sets	0	0
9	Simple Negation	0	0
10	Indefinite Knowledge	0	0
11	Basic coreference	0	0
12	Conjunction	0	0
13	Compound coreference	0	0
14	Time manipulation	0.23	0.2
15	Basic deduction	0	0
16	Basic induction	42.75	45.3
17	Positional reasoning	3.46	4.2
18	Reasoning about size	0.67	2.1