Making a chatbot is really hard, because you need a lot of data and it needs to be of good quality. You cannot use a random chat message log and train your model, it will usually have way too much noise. Instead I highly recommend that you build it from scratch. Then you will start getting great results. You also have to use a proper spell checking algorithm and understand the basics of how a full text search engine actually works and know the purpose of using stop words and synonyms.
I would love to present our virtual assistant "James" to the rest of the world, but we have only recently started to build english training data. Currently it only understands english with the use of synonyms, so it isn't working as well as it should.
If you're interested in trying it out you can play with it on the webpage of one of the largest banks in Norway http://srbank.no, the chat icon is at the bottom right. You can ask questions in English, but it will unfortunately only answer in Norwegian at this time, so you have to use google translate too see if it answered correctly. "James" do support language detection though, but we have just recently starting to add the english answers. In a few weeks and James will answer in English :-)
Currently James can predict over 1500 intents related to banking and insurance. Here are some of the more complex questions James can handle:
Want to open a savings account for my daughter
Can I adjust the limit on my sons bank card
What can you tell me about my pension
Do you have a pension calculator?
I was out drinking last night and lost my credit card
If anyone says that chatbots aren't working and will never work, I can say this with real data to back me up, it works very well! The feedback we have gotten the last few months has been amazing! And I'm 100% sure that chatbots will become a big industry the next years. Complex UI's will finally be good riddance.
Feel free to ask me anything if you have any questions.
Yes, all the data we use to train our model on, is handcrafted by our AI trainers. What they do is figure out all the different questions that are most common to ask for each intent we want to predict. We have built a tool that helps them creating all these sentences, often we use templates with keywords.
It is still under development, we started at the end of last summer. We use Torch and Python-NLTK.
Why have you chosen torch instead of staying completely in python? I assume that you use nltk for tagging, entity regoncition, ... and torch for your classifier.
What kind of model is possible in LUA-torch that could not be done in python with caffe/tensorflow?
The bot looks good. Have you tried AWS Lex? It has a pretty good English language model (same as Alexa) - you just add a bunch of example sentences per intent, and it does the matching
Actually no ;-) We do use Amazon Polly though, and Google's Speech to Text. Works very well with our chatbot, but none of our clients want stt and tts. Text is what they all want. Maybe later when both solutions become more reliable for nordic languages. English is very good though, on both sides.
I would love to present our virtual assistant "James" to the rest of the world, but we have only recently started to build english training data. Currently it only understands english with the use of synonyms, so it isn't working as well as it should.
If you're interested in trying it out you can play with it on the webpage of one of the largest banks in Norway http://srbank.no, the chat icon is at the bottom right. You can ask questions in English, but it will unfortunately only answer in Norwegian at this time, so you have to use google translate too see if it answered correctly. "James" do support language detection though, but we have just recently starting to add the english answers. In a few weeks and James will answer in English :-)
Currently James can predict over 1500 intents related to banking and insurance. Here are some of the more complex questions James can handle:
Want to open a savings account for my daughter
Can I adjust the limit on my sons bank card
What can you tell me about my pension
Do you have a pension calculator?
I was out drinking last night and lost my credit card
If anyone says that chatbots aren't working and will never work, I can say this with real data to back me up, it works very well! The feedback we have gotten the last few months has been amazing! And I'm 100% sure that chatbots will become a big industry the next years. Complex UI's will finally be good riddance.
Feel free to ask me anything if you have any questions.