About

This repository implements a model to do style transfer for a chatbot which is reponding to a conversation.

---

Model

• Inputs to the model: Conversation History (left), Generic Response (centre)

• Output: Courteous Response (right).

• The Conversation History is encoded by hierarchical BiLSTM to a Conversational Context vector c.

  • First Level Bi-LSTM encodes utterances
  • Utterance representations are concatenated to their respective DeepMoji representations
  • Second Level Bi-LSTM encodes the sequence of utterances

• The encoder encodes the Generic Response into hidden states h_i .

• Response tokens are decoded one at a time. Attention α_i , and vocabulary distributions (p_vocab ) are computed, and combined using p_gen to produce output distribution.

---

This code has been built over the code of the ACL 2017 paper Get To The Point: Summarization with Pointer-Generator Networks. This code is currently not being maintained.

---

run_summarization.py is the top-level file to train, evaluate or test models