https://github.com/s1998/all-but-the-top

https://github.com/s1998/all-but-the-top

abusive-language davidson deep-learning embeddings f1-acc hate-speech iclr iclr2018 natural-language-processing word-embeddings

Last synced: 6 months ago
JSON representation

• Host: GitHub
• URL: https://github.com/s1998/all-but-the-top
• Owner: s1998
• Created: 2019-09-29T23:44:14.000Z (about 6 years ago)
• Default Branch: master
• Last Pushed: 2019-11-26T15:14:38.000Z (almost 6 years ago)
• Last Synced: 2025-02-01T18:27:53.349Z (8 months ago)
• Topics: abusive-language, davidson, deep-learning, embeddings, f1-acc, hate-speech, iclr, iclr2018, natural-language-processing, word-embeddings
• Language: Python
• Size: 80.1 KB
• Stars: 2
• Watchers: 3
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          

# All-but-the-top

Implementation of the paper [All-but-the-top](https://openreview.net/forum?id=HkuGJ3kCb) from ICLR 2018.

## Instructions to use

To run, use the file runner.py

Libraries used: Keras with tensorflow backend.

### Sample plot

![image-of-principal-components](https://raw.githubusercontent.com/s1998/All-but-the-top/master/images/gloveFreqPLot.png)

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)