We are big fans of science and would be happy to make it more popular among people. Originally we wanted to build a service that summarizes and simplifies scientific articles so they become easier to understand and this summaries could be used in social media. But such project is far more difficult and model training requires lots of time.
So we decided to start with a simpler service: telegram bot which offers headline for your text.
What it does
- Give our bot some text and it will create the best headline it can come up with.
- It will also do some text stylization. This style transfer mechanism will be a basis for article simplification in our future development.
Challenges we ran into
We experienced two problems with AWS: account verification took long time and acquiring a GPU instance surprisingly became a problem, so this time we got along without it.
What we used
During hackathon we experimented with:
- tensor2tensor library (got poor rouge score);
- tensorflow-based RNN with attention.
What's next for TLDR bot
If we stick to our original plan and continue working on it our bot will learn to read scientific articles and paraphrase them in such words that your 5yo kids or grandparents could understand.