/Structured-Self-Attention

A Structured Self-attentive Sentence Embedding

Primary LanguagePythonMIT LicenseMIT

Structured Self-attentive sentence embeddings

Implementation for the paper A Structured Self-Attentive Sentence Embedding, which was published in ICLR 2017: https://arxiv.org/abs/1703.03130 .

USAGE:

For binary sentiment classification on imdb dataset run : python classification.py "binary"

For multiclass classification on reuters dataset run : python classification.py "multiclass"

You can change the model parameters in the model_params.json file Other tranining parameters like number of attention hops etc can be configured in the config.json file.

If you want to use pretrained glove embeddings , set the use_embeddings parameter to "True" ,default is set to False. Do not forget to download the glove.6B.50d.txt and place it in the glove folder.

Implemented:

  • Classification using self attention
  • Regularization using Frobenius norm
  • Gradient clipping
  • Visualizing the attention weights

Instead of pruning ,used averaging over the sentence embeddings.

Visualization:

After training, the model is tested on 100 test points. Attention weights for the 100 test data are retrieved and used to visualize over the text using heatmaps. A file visualization.html gets saved in the visualization/ folder after successful training. The visualization code was provided by Zhouhan Lin (@hantek). Many thanks.

Below is a shot of the visualization on few datapoints. alt text

Training accuracy 93.4% Tested on 1000 points with 90.2% accuracy