A bot with a soul
Natural language processing is possibly the hardest problem for machines and yet so natural and easy for humans. Techniques based on deep learning proved their efficiency on a number of NLP tasks, challenging us to build the most powerful systems yet. We took the challenge of understanding and generating a popular genre of poetry know as "kafanske pesme" - a highly emotional form of entertainment in Balkans.
To measure and improve the quality of generated music, users can click on the verse they don't like and the new one will be generated instead. We use that addition training data to improve the quality of the system.
The bot, named kafanabot
for obviuous reasons, was built using Tensorflow
and was trained on the dataset of 30MB of most popular songs for two days on a single GPU.
A simplified version of the bot was used as a case study in our deep learning class
and can be found here: https://github.com/wecliqued/deep_learning/tree/master/seq2seq