Generating theatre plays using techniques from Neural Machine Translation
Saad Obaid ul Islam
2 of 6
Goal
An AI that can generate a theatre play given a scenario.
3 of 6
Transformer
4 of 6
Techniques
De-summarization
Use a pre-trained transformer trained on summarization data and fine-tune it on scenario - script data. (flip the inputs on encoder and decoder).
Challenges: No literature available. No metrics to measure de-summarization.
Train our own encoder-decoder architecture on scenario-script data.
Challenges: Lots of data and computing resources required.
Script has several variable like dialog, setting of the scene.
5 of 6
Progress
Testing de-summarization technique using summarization data by flipping the inputs of encoder and decoder.
Output looks good.
Problem: Will this work for script data?
Input: Largest Czech COVID vaccination centre open but waiting for shots.
["Thousands of people in the Czech Republic have been queuing for hours to get vaccinated against the deadly virus known as CO, which is believed to have killed more than 100 people in the past few weeks, as the country's first CO vaccination centre opened in the capital Prague on Friday, as the country's first CO vaccination centre opened its first in the country, as the centre opened its second centre in the capital city Prague, as the country's first CO vaccination centre opened its second centre in the capital Prague, as the country's first CO vaccination centre opened."]