Build a Chatbot - ML for Hackers #6

Key Takeaways

open the pod bay doors Hal I'm afraid I can't do that s why not Hal because I'm overfitted on the wrong data what the hello world welcome to serology in today's episode we're going to learn how to build a chatbot chat Bots have come a long way in the past few years remember the smarter child bot on AIM that thing was pretty fun at the time but now it's like do you even AI bro the future of software is one where Bots will slowly replace our need to fiddle with clunky UI we'll be able to just ask our AI to book an Uber or find the best taco place on Yelp for US service layers will be hidden under a plain English conversational layer when I think of real AI I think of a human level chatbot the OG computer scientist Alan Turing proposed a test to judge whether or not a machine exhibited human level intelligence by having the human observe a conversation between a human and a machine if it couldn't tell if the machine was a human or not it passed the test so far no chatbot has passed a Turing test but we'll get there traditionally chatbots have used a retrieval-based model to communicate in a retrieval based model programmers code in a set of predefined responses and some kind of heuristic to pick the appropriate response based on the input and context the first chat boots were just rule-based expression matching like if I ask the exact phrase will I ever get laid it responds no every time but more recently companies have started using more complex heuristics like using a machine learning classifier Facebook messenger's chatbot API is an example of this you can hardcode responses to potential questions and the system classifies words to understand intent so you could either ask what day is it today or today is what day and it would understand that both questions although worded differently have the same intent the harder chat up model is generative these don't rely on any predefined responses whatsoever they generate them from scratch two Google researchers released a paper called a neural conversational model where they train a neural net on two data sets to do this first on a movie dialogue data set so it would be able to speak conversational English then on an IT support data set so it had domain knowledge when they tested it on a real human asking for support it was remarkably efficient at helping them solve their problem without any hard-coded responses just by giving it data and training it okay so what kind of Bot do we want to build well when building a chat bot we have to think about possible constraints are we operating on a closed domain or an open domain in an open domain the conversation can go anywhere there are an infinite number of things to talk about in a closed domain the conversation focuses on a single subject if we want to operate on an open domain using a generative model that's pretty much AGI so we're not quite there yet if we use an open domain with a retrieval model we'd have to hardcode literally everything so also impossible so right now we can build a chatbot in a closed domain using either retrieval or generative model okay let's add in one more constraint do we want it to have long or short conversations short conversations are easy you just output a single response to a single question long conversations are a bit harder the AI has to keep track of what's being said that is the context over a series of questions from the user support topics would be a good example of this we could go the easy route and use a retrieval model if all we want is a bot to give us the local weather but if we want our bot to have a long conversation with us about the weather like what's the weather in SF is my family safe where can I find a new family then we should go for a generative model we need lots of data to train our bot on a generative model like a big chat log or a knowledge base and when done well that's pretty much the bleeding edge which means that's what we have to do do so we're going to recreate the results from the neural conversational model paper using the Deep learning library torch in the Lua programming language let's collect our data set first we'll be using the Cornell movie dialogue data set and we'll set our variables from the command line to how much of the data set we want to use and the minimum frequency of words that we keep in our vocabulary our next step is to build a model we use our command line arguments to help determine the size of the model the two variables being the number of hidden layers and the word count of our data set in our case this will be a sequence to sequence Model A sequence to sequence model cons cons of two long short-term memory recurrent neural networks the first neural net is an encoder it processes the input the second neural net is the decoder and it generates the output so why the sequence to sequence model yes deep neural Nets are awesome but they require the dimensionality of the inputs and outputs to be a fix size we're accepting a sequence of words in a sentence and outputting a new sequence of words so we need a sequence learning model that can learn on data with long range memory dependencies lstm architecture is the Natural Choice the encoder lstm turns the input sentence of variable length into to a fix dimensional Vector representation we can think of this as the thought Vector so given a large enough data set of questions and responses it will recognize the closeness of a set of questions and represent them as a single thought Vector what time is it what's the time yoyo what's the time Isle myle will all fall into a single thought Vector so after training we'll have a huge set of not just synapse weights but thought vectors as well next we'll want to add in some hyperparameters we want to use a class nll Criterion for our model NL stands for negative log likelihood this will help us obtain log probabilities from our input data which will help us improve our sentence predictions the learning rate and momentum helps Pace our time steps and Decay factor and min mean error help improve our learning rate while training then we'll make sure Cuda is enabled and start training our model using back propagation in each Epoch or run we'll declare our error and timer variables and loop through each example in each batch the default batch size is a thousand examples for each of those examples we'll get the input sentence and the target sentence we'll use the input and the target as parameters to train our model then we'll want to error check and make sure we record our progress at the end of each iteration we save our model if it improves and update the learning rate boom that's it after training this baby on AWS we can have a conversation with it the more data you give it the better it's going to get and if you're going to do this add a filter for curse words I'm looking at you Microsoft this will eventually automate a lot of support jobs away completely so if there are any government people in the house right now let's get on that basic income Jam ASAP unless you want a revolution for more info check out the links down below and please subscribe for more ml videos for now I've got to go fix a memory leak so thanks for watching