A Knowledge-Grounded Neural Conversation Model
MSR, AAAI 2018
TL;DR
- 過於 general
-
沒有提供有用資訊
解決 chatbot 回覆
的問題
Example
User input: Going to Kusakabe tonight.
Neural model: Have a great time!
Human: You’ll love it! Try omasake, the best in town.
對話生成流程
- identify focus in input (by keyword matching, NER, etc.)
- retrieve relevant facts of the focus.
- feed (1) conversation and (2) relevant facts into NN.
Model
Dialogue Encoder / Decoder
Seq2Seq: GRU-RNN
Facts Encoder
Memory Network Sukhbaatar et al. 2015
Facts Encoder(Cont'd)
Memory Network - RNN
Experiment
Data
Knowledge Foursquare
Conversation
- General Dataset 23M Twitter 3-turn conv.
- Grounded Dataset 1M [Twitter 2-turn conv. with @handle] augmented w/ Foursquare
Tasks (Multi-task learning)
- NoFact w/o fact encoder
- Fact train on
- AutoEncoder train on
Multi-Task Learning (MTL)
Hard param sharing
Soft param sharing
Tasks (Multi-task learning)
- NoFact w/o fact encoder
- Fact train on
- AutoEncoder train on
Systems
- Seq2Seq NoFact-General
- MTask NoFact-General + NoFact-Grounded
- MTask-R NoFact-General + Fact-Grounded
- MTask-F NoFact-General + Autoencoder-Grounded
- MTask-RF NoFact-General + Fact-Grounded + Autoencoder-Grounded
Perplexity
BLEU-4 & Div.
Human Eval.
Example
A knowledge-grounded neural conversation model
By qitar888
A knowledge-grounded neural conversation model
the lab report of a paper published by MSR @ AAAI '18
