一 Datasets
1. ubuntu chat corpus: http://daviduthus.org/UCC/
2. dialogue datasets (circa 2005)
二 Tasks
task1: next utterance classfication
the dual encoder model
task2: large copurs next-utterance retrievl
task3: narural language generation
hierachical encoder-decoder architecture(bengio AAAI 2015)
三 Interesting topics
(Interesting! ) exploring ethical and social challenge:
bias:
bias in image classification, debiasing word embeddings(bolukbasi et al 2016), hate and offensive language detection (Davidson at al 2017)
privacy
safety
value alignment
adversarial examples
四 Challenges
Challenge: how to evaluation ?
evaluation for generative models:
perplexity
word error rate
word overlap metrics, BLEU
camparing response with human
五 adversarial examples