[ref][tutorial] Deep Learning for Chatbots, Part 2 – Implementing a Retrieval-Based Model in Tensorflow

[at15]

http://www.wildml.com/2016/07/deep-learning-for-chatbots-2-retrieval-based-model-tensorflow/

• Dual Encoder LSTM network (translation is encoder, decoder)
• [ ] the recall k evaluation, need to pick 10 response, (1 is correct, other 9 random), then pick first k with biggest possibility, but how to get the 10 response in real world, the bot can run through the entire response set to get the biggest possibility when data is large
  • [ ] the blog mentioned google use clustering Smart Reply: Automated Response Suggestion for Email
• [x] random baseline
• [x] tfidf
• [ ] seq2seq in tf
• [ ] seq2seq in mxnet

Text processing

• using builtin tf.contrib.learn.preprocessing.VocabularyProcessor for words to id

TF

• [ ] it uses the estimator API (instead of manually call sess.run)

[at15]

I think will switch to example that use smaller dataset (such as Cornell movie) for now and come back to this one later, it's hard to play with it when the data size is big