Bessie

• 

  Example script generated

• 

  Front-end

Inspiration

Terminal nerd-ism and Holly.

What it does

Generates classic doctor who scripts using a neural net.

How we built it

Used Keras to build a model, Beautiful soup to download and process scripts. Website front-end with html/css.

Challenges we ran into

"doctor: well stepple the doctor"

Text formatting was wack. Tuning the model was hard. Integrating back-end and front-end needed. Unfortunately we didn't have time to perfect the model nor integrate the front end website with the model. Furthermore, there were numerous problems encountered while training mainly because 32GiB of RAM was apparently not enough for 12 million characters and 40 years of doctor who. Some expertise with google cloud would've been helpful here. Or a time machine.

Accomplishments that we're proud of

Surprisingly, the model learned a lot of the structure of scripts almost flawlessly. This includes things like newlines between characters, placing colons before the name of characters that are speaking, and even stage directions!

We had some cool ideas to make the model output more natural text that we implemented like text find and replace using edit distances to match doctor who key words that are just slightly misspelled.

What we learned

Computer no like training big model. Computer die. Back-ends hard.

What's next for Bessie

Integrate back-end and front-end, use more scripts to generate interesting things e.g. the thick of it and doctor who for a sweary 12.

Built With

• beautiful-soup
• css
• html
• keras
• python
• tensorflow