Don’t expect accurate answers here - it’s an experimental weekend hack.
With this experiment, we wanted to understand how we could train a neural network on new external datasets very quickly and study the difference in the language and answers depending on the topic of the dataset.
No scripted chatbot here, we are talking of neural-nets trained in the wild :) We are using the celebrated seq-2-seq model which computes a “thought-vector” from an input sentence and generates an output sentence conditioned on this vector. We gathered various datasets from stackexchange and launched a big overnight training of our models to have some surreal morning coffee talks with our AI.
There’s still a ton of work for the answers to start to make sense (longer training and bigger datasets would improve the quality of course, we can also easily add components to improve the variety and coherence of the responses) but a difference can definitely be noticed based on the dataset subset.
This will allow us to test a variety of new datasets way faster than before.
Let us know what you think, and if you think of a cool dataset to test, just let us know on Github.
Do you have any details on your experience training the models? Was this using vanilla RNNs, LSTM cells, did you use an attention mechanism, etc.? How long did you train for?
Right now, it doesn't work at all. Would be great to test and iterate on it a bit and to have some sense to the answers before exposing it to the world.
11 comments
[ 3.8 ms ] story [ 81.4 ms ] threadDon’t expect accurate answers here - it’s an experimental weekend hack.
With this experiment, we wanted to understand how we could train a neural network on new external datasets very quickly and study the difference in the language and answers depending on the topic of the dataset.
No scripted chatbot here, we are talking of neural-nets trained in the wild :) We are using the celebrated seq-2-seq model which computes a “thought-vector” from an input sentence and generates an output sentence conditioned on this vector. We gathered various datasets from stackexchange and launched a big overnight training of our models to have some surreal morning coffee talks with our AI.
There’s still a ton of work for the answers to start to make sense (longer training and bigger datasets would improve the quality of course, we can also easily add components to improve the variety and coherence of the responses) but a difference can definitely be noticed based on the dataset subset.
This will allow us to test a variety of new datasets way faster than before. Let us know what you think, and if you think of a cool dataset to test, just let us know on Github.
me: yes
gg: I'll send screen shots
me: yes
gg: I'll send screen shots
me: yes
gg: I'll send screen shots
me: yes
gg: I'll send screen shots
....
http://imgur.com/kLtsooB
Full Stack Dev: Do n't do this
Me: Yes
Them: I'll update the tag
Me: Ok thank you
Them: The second one then.