Here are a few that might be worth looking at: NVidia NeMo https://github.com/NVIDIA/NeMo Microsoft Icecaps https://github.com/microsoft/icecaps … Google simulated dialog https://github.com/google-research-datasets/simulated-dialogue … DREAM (based on BERT) https://github.com/nlpdata/dream Facebook ParlAIhttps://github.com/facebookresearch/ParlAI …
ML community: have anything better than GPT-2 for conversation? anything that can track discourse as it evolves over conversation? if not, any chance you might need significant new ideas rather than just more data & GPU? impossible that symbolic representation would help?https://twitter.com/powooli/status/1194262078367903744 …
-
-
-
thanks for the links!
End of conversation
New conversation -
-
-
Are you familiar with Microsoft's DialoGPT system; it builds off GPT approach, but is specifically geared towards conversational dialog. https://arxiv.org/pdf/1911.00536.pdf …
-
thanks and pity they don’t have an instance ready to try.
End of conversation
New conversation -
-
-
I've been tracking the problem for a long time and I can say with some confidence that such thing does not exists. The issues encountered on GPT2 are not new, exactly the same happens with simple Seq2Seq models. It's just more evident here since is much better on grammar.
-
So yes! ML desperately needs some way to integrate and manipulate symbols. Otherwise models can't remember or reason. Something that we can call "the missing link of AI".
- 1 more reply
New conversation -
-
-
We
@hrilab presented a paper at AAAI-2019 about how discourse must be tracked and reasoned with for coreference resolution (https://hrilab.tufts.edu/publications/sarathy2019aaai.pdf …), and how it can be done computationally. Not quite "conversation", but we are interested in situated dialogue and discourseThanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
SIGDIAL community in NLP have been working on tracking dialogue context as it evolves in conversation. Both symbolic and ML methods used https://www.aclweb.org/anthology/venues/sigdial/ …
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Significant new ideas required? Sure. Symbolic representations? Yes, why not. But then the hard part. What should these symbols be? How do you cause cognition to emerge? No conversational skills without cognition; and cognition won't "just emerge" in a Pytorch/TF neural net.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Interesting observations, Gary! At
@IBMResearch we have been recently exploring some intriguing combination of planning+NLP techniques to model sophisticated goal-directed conversations
http://arxiv.org/abs/1910.08137 Thought you might be interested :)https://twitter.com/cjmuise/status/1186247146036781058 …Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.