Implementation of 1D Convolutional Neural Network for Sentence Classification based on the paper by Yoon Kim(2014)
- Large Movie Review Dataset (https://ai.stanford.edu/~amaas/data/sentiment/)
- TREC (http://cogcomp.cs.illinois.edu/Data/QA/QC/)
- 100-dimensional GloVe vectors are used as embeddings.
- 2 Embedding layers are used, one with static weights and the other with dynamic weights which can be fine-tuned.
- Kernel size for Convolutions - {3, 4, 5} with 100 filters each.
- Each Embedding layer is connected to a different CNN block.
- All outputs from the CNN blocks are concatenated and given to the fully connected layer.
- Dropout rate set to 0.5
- Batch size set to 50 and model trained for 6 epochs.
*CNNPool in the diagram below refers to a set of 3 CNNBlocks. Each CNNBlock consists of a 1D Convolutional layer follwed by a Global Max Pooling Layer. All 3 CNNBlocks have different kernel sizes and run in parallel.
