Data are loading

Explanations By Boundary Exploration for Textual data


Choose a neural network (NN) and a dataset size

GPT-2
Transformer NN with attention and no reconstruction task
This implementation uses the huggingface.co implementation for sequence classification of GPT-2. Texts are embedded in 1,024 dimensional space.

10,000 entries 50,000 entries 100,000 entries
SSA
Recurrent NN (RNN) with attention and no reconstruction task

This implementation uses a structured self-attentive sentence embedding classifier implemented here. Texts are embedded in 50 dimensional space.

10,000 entries 50,000 entries 100,000 entries
Reconstruction RNN
RNN without attention and with a reconstruction task

This implementation uses a simple recurrent neural network presented here. Texts are embedded in 1,024 dimensional space. Texts are troncated at the 20-th word. For this network and this dataset UMAP is used with a number of neighbors set at three instead of two.

10,000 entries

EBBE-Text is here presented with 3 differents NN. However, EBBE-Text could be use with any classifier which create an embedding for each inputed data.