DSTC7 challenge
Basic codebase from : https://github.com/IBM/dstc7-noesis Please refer https://github.com/IBM/dstc7-noesis/tree/master/noesis-tf for requirements and training data format.
- BiDirectional LSTM (arg: --bidirectional)
- Mean and Max-pool based feature encoding (arg: --feature_type=mean/max)
- CNN based max-pooling of features (arg: --feature_type cnn)
- Attention (arg: --attention)
- Fast GRNN (arg: --fastgrnn)
- Using C-DSSM based sentence (pass on in input data) embeddings (arg: --dssm)
- Custom Word2Vec embedding learning on dataset (For custom word2vec training, read here
- Matrix factorization to reduce time complexity and prevent overfitting (arg: --factorization)
- Extend training set by taking 7th turn onwards as positive utterances