3 ms·
I'm working for a startup called Boost.ai where we have decided to build our own deep learning model and create our own training data. We use several models tha
by olavgg 10y ago
I'm working for a startup called Boost.ai where we have decided to build our own deep learning model and create our own training data. We use several models that predicts what word comes next, take care of mispellings and dialects. We also have memory so it can remember contexts when you ask a new question or it predicts that you want to start with a new context.
If you want something kick ass, I highly recommend building it yourself. The bar set by IBM's Watson, Nuance's Nina or IPSoft's Amelia isn't actually very high. For non english languages, anyone who has some knowledge about NLP and deep learning will easily surpass them.
- choxi 10y agoAny pointers on where to get started? I've messed with Karpathy's RNN example using Shakespeare and PG essays, but don't know where to go from there.
- olavgg 10y agoKarpathy's example tries to predict the next character, there is a fork of it that tries to predict the next word. https://github.com/larspars/word-rnn https://github.com/larspars/word-rnn RNN is great approach if you want to play around with text generation with deep learning. But for a chatbot, deep learning alone is not there yet. We create our own intents based on the domain and predict the intent of each question. We have also made it looks smarter by creating an intent hierarchy where we try to do multiple predictions for a question with a goal to drill the question down in the intent tree. In that way we know that the question is about a bank card and can figure out if you want a new one, block it, increase credit, set limits and so on.
- chiefalchemist 10y agoRe: "If you want something kick ass, I highly recommend building it yourself." That might be true, but the ability to spin something up without all that overhead is any MVP'ers wet dream, yes?