Chatbot.md

Chatbot

Overview

• Non-Task-Oriented Dialogue System
  • Retrieval-base Method
  • Generation-base Method

Retrieval-Based Chatbot

Response Selction

Roadmap

• Message-Response Matching for Single-Turn Response Selection
  • Framework
    • Matching with sentance embeddings --extension--> Matching with external knowledge (Topic Aware Attentive RNN)
    • Matching with message-response interaction --extension--> Matching with multiple levels of representations (Knowledge Enhanced Hybrid NN)
      • Similarity matrix-based interaction
      • Attention-based interaction
  • Insights from empirical studies
  • Extension: Matching with external knowledge
• Context-Response Matching for Multi-Turn Response Selection
  • Framework
    • Embedding --> Matching
      • Dual-LSTM
      • Multi-view Response Selection Model
      • Deep Learning to Respond (DL2R)
    • Representation --> Matching --> Aggregation
      • Sequential Matching Network (SMN)
      • Sequential Attention Network (SAN)
  • Insights from empirical studies
• Other emerging research topics
  • Matching with better representations
  • Matching with unlabeled data

Generation-Based Chatbots

• Sequence-to-sequence
  • Neural responding machine
  • Encoding-decoding framework
  • Model variants with different attention
• Attention mechanism
• Bi-directional modeling

Datasets for Empirical Studies

• Ubuntu Dialogue Corpus - 1.0, 2.0, paper, kaggle

• Douban Conversation Corpus - github

Evaluation Metrics

$R_n @k$: For each message, if the only positive response is ranked within top $k$ position of $n$ candidates, then $R_n @k = 1$. The final result is the average on messages.

Resources

Article

• 知乎 - 小哥哥，檢索式chatbot瞭解一下？
• 檢索式人工智障識記