Skip to content

Latest commit

 

History

History
51 lines (35 loc) · 2.15 KB

README.md

File metadata and controls

51 lines (35 loc) · 2.15 KB
Raw

DeepGuru AI

DeepGuru is a twitter bot generating inspirational quotes. It uses deep learning to generate tweets.

GIF showing the widget

You can check it here: https://twitter.com/DeepGuruAI

Technology

The bot uses the ULMFiT model.

It is trained on a corpus of tweets created for this task (2 corpora of ~4k tweets). Because the ULMFiT model is pre-trained on an extract of Wikipedia, it is able to train well with a small dataset (~4k tweets for the first version).

How it works

The process is the following:

  • Create a dataset with tweets of the target style (in DeepGuru case, philosophical and zen tweets)
  • Fine-train the ULMFiT english model on the corpus (Notebook "Model Training")
  • Use the trained model to generate tweets (script "QuoteGenerator.py")
  • Selection step: Filter the generated tweets (Notebook "Post-processing of generated quotes")
    • At the moment, there is a manual step of filtering the generated tweets. The tweets are not modified but around 30% are rejected
    • I am working on a second neural network to automatically select the best generated tweets
  • Twitter bot: Tweet a few times a day using the generated data, with a cron task (script "bot.py")

Next steps

  • Write a technical post with an in-depth explanation of the approach and technical parts
  • Finalize the second neural network to automatically select the best generated tweets
  • Experiment with data: new corpus and mixing of corpus
  • Experiment with the generation algorithm
  • Experiment with model: test other models like google BERT

Dependencies

The project uses:

To use it, you have to install:

  • Anaconda distribution with Python 3
  • Fastai (which is based on Pytorch)
  • Tweepy
  • LabelMyTextWidget