Adding to @Matthew & @ Mariya’s conversation, I thought this talk from Percy Liang was pretty interesting
There’s a lot of preface, but the experiment was interesting. Based on the idea that language is based on the need to motivate an action in the world, Stanford NLP researchers built a game where players (mechanical turk workers) have to come up with language to move virtual blocks to recreate a given pattern. The game does not have language, so players train the game using their own language–whether English, Polish, or even Polish Notation). The game aggregates the common words given by the players to create a working body of language.
Thanks for sharing, @melissa.fabros! I love your line too about how “language needs to motivate an action in the world.” I think if we all kept that in mind, we would be much better communicators ourselves.
BTW - for anyone who likes watching videos, here’s the YouTube link to Percy’s talk on “Natural Language Understand: Foundations & State Of The Art”: https://www.youtube.com/watch?v=mhHfnhh-pB4
One paper I can’t stop reading is the one Jeremy touched on just after the chatbot section of the lecture regarding “Recurrent Entity Networks”. I think they will be impactful.
Also, if viv.ai is the real deal, the data/exposure they will get from ALL Samsung devices(they acquired them last year) may make them a real force in this market. They plan on opening up their tech so everyone can build on the assistant. Which, is usually a winning strategy as the App store/Facebook proved.
Here’s a TensorFlow implementation of the Recurrent Entity Networks paper:
In my discussions with the developer, he was of the belief that they’re only currently good for solving toy problems, similar to the end-to-end memory networks we examined in class.
Your thoughts? Chatbots still doomed to be dumb?
cc: @jeremy@rachel
Maybe it’s just my imagination(or perception of the paper), but I can see how they(or a future variation of them) could be used to learn and track parts of our lives.
Chatbots and AI assistants are just getting started imo. The majority have not been implemented well. But some are proving useful. https://x.ai/ comes to mind.