项目作者: iamgroot42
项目描述 :
Code-base for our submission to the DSTC7 challenge (Subtask-1)
高级语言: C#
项目地址: git://github.com/iamgroot42/dstc7.git
dstc7
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.
Additional features
- 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