All-but-the-top

Implementation of the paper All-but-the-top from ICLR 2018.

Instructions to use

To run, use the file runner.py

Libraries used: Keras with tensorflow backend.

Sample plot

Two principal components of the embeddings with color map for frequency.

Results

Results obtained on Davidson et al (2017) using two different embeddings.

Model	Preprocessed				Postprocessed
	P	R	F1	Acc	P	R	F1	Acc
AvgPool	0.811	0.721	0.756	0.762	0.855	0.887	0.862	0.887
MaxPool	0.779	0.834	0.792	0.787	0.888	0.903	0.884	0.903
CNN	0.885	0.903	0.880	0.903	0.890	0.905	0.892	0.905
GRU	0.894	0.907	0.898	0.907	0.899	0.914	0.902	0.914

Effects of using post processing on Glove Embeddings on Davidson et Al(2017)

Model	Preprocessed				Postprocessed
	P	R	F1	Acc	P	R	F1	Acc
AvgPool	0.787	0.732	0.754	0.782	0.898	0.893	0.868	0.883
MaxPool	0.703	0.756	0.724	0.751	0.891	0.887	0.872	0.887
CNN	0.838	0.887	0.861	0.887	0.875	0.893	0.875	0.891
GRU	0.854	0.904	0.878	0.904	0.910	0.903	0.881	0.903

Effects of using post processing on Word2Vec Embeddings on Davidson et Al(2017)