Alexa

References

Brainstorm 2/20/17

Questions:

1. What are the kinds of utterances that people give when talking about a topic? How can we represent their meaning? What is a rough classification or grammar?
2. Simulate conversations about a popular topic.
3. What makes a conversation interesting?

Resources:

1. Semantic parsing. Data gathering pipeline.
2. What are other groups doing?
3. List of conversation topics. See pixie-other/

Blogs, writeups, code

• WildML chatbot
• Attention and memory

OLD

• Dialogue
  • [LMRG16] Deep reinforcement learning for dialogue generation paper
  • [DGZB16] EVALUATING PREREQUISITE QUALITIES FOR LEARNING END-TO-END DIALOG SYSTEMS paper
  • [VL15] A Neural Conversational Model paper
• Neural nets
  • [SSWF15] End-To-End Memory Networks
    • Architecture
      • Input \(x_i\)
      • Convert input into memory \(m_i = A x_i\), \(A\in \R^{d\times V}\)
      • Output vecor \(c_i = C x_i\).
      • Query \(q\)
      • Embedded query \(u=Bq\).
      • Match between queries and memory \(p_i = \text{softmax}(u^Tm_i)\).
      • Response is weighted sum \(o = \sum_i p_ic_i\).
      • Predicted answer \(\wh a = \text{softmax}(W(o+u))\).
    • For multiple layers, \(u^{k+1}=u^k + o^k\). Each layer has \(A^k,C^k\) to embed inputs.
      • Two types of weight tying
        • output = input above, \(A^{k+1}=C^k\).
        • RNN: \(A^k =A, C^k=C\), and modify \(u^{k+1} = H u^k + o^k\).
• OpenSubtitles
  • Main page
  • ?

Questions

How does the beam search actually work? Do we just have \(\Pj(w_1)\Pj(w_2|w_1)\Pj(w_3|w_2,w_1)\) or is there some more complicated energy-based model? (Don’t normalize?)

How to promote consistency in answers? Hidden state.