BBC News Content Classification is a multiclass classification problem. The dataset can be found at this source : http://mlg.ucd.ie/datasets/bbc.html
A simple Embedding from scratch and LSTM based Deep Neural Network is used to train and test the network. The model architecture is as follows:
| Layer(Type) | Output Shape | Parameters |
|---|---|---|
| embedding (Embedding) | (None, 120, 64) | 512000 |
| lstm (LSTM) | (None, 64) | 33024 |
| dense (Dense) | (None, 256) | 16640 |
| dropout (Dropout) | (None, 256) | 0 |
| dense_1 (Dense) | (None, 6) | 16640 |
| Total params: 563,206 | ||
| Trainable params: 563,206 |
The embedding of 8000 different words, each embedded with a seperate vector having 64 dimensions. The embedding are initialized as random and are trained along with the training process. The trained embedding can be interacted in depth in the following URL: https://projector.tensorflow.org/?config=https://gist.githubusercontent.com/Arko98/b857c056350b0c38804de25b0e51ef03/raw/bcee757847d369accbfc0adc6bff7b1c603b2b76/projector_config.json
The Model obtained results as follows:
- Training Loss: 8.6603e-05
- Training Accuracy: 100%
- Testing Loss: 0.3070
- Testing Accuracy: 94.83%
