Final Project: Chat with Woolf

Working link as follows:

For my final project, I decided to continue my nanogenmo project, except I would make it more interactive.

I had to revisit the material from class, in which I wanted to reread the definition of Computational Narrative Intelligence before proceeding with my idea.

Narrative intelligence is the ability to craft, tell, understand, and respond affectively to stories.

One of the Medium posts that was linked on the syllabus, poised this question:

One open challenge in interactive narrative is non-programmer authorial intent. How do we make it possible for non-technical storytellers to instill their intent, goals, and beliefs about good story experiences?

I wanted to tackle the question with my final project, such that if non-programmers such as myself, can interact with a narrative, perhaps from a novel and have people craft, respond to the generative gestation of the story. I thought of making the appearance of a chatbot from my trained model of Woolf and Friends. I remember watching the movie, Her and there was a scene about whether AIs can learn from others and Alan Watts was in the scene. I was inspired by the idea that we can chat to an influential figure from the past.

Scene where AIs talk to each other
Movie HER
  1. I believe in expressive choice: narrative choice doesn’t have consequences and makes the players feel like they’re in control and they can be themselves.



  1. Collect a large corpus of text.
  2. Train character level LSTM network
  3. Once training is done, sample data input
  4. Style the CSS sheet
  5. Import Speech Library

Things I don’t know how to do

How to make the conversation more dynamic so that when you’re asking a question it picks up the phrase and asks a question back?

Learnings of what didn’t work

  • since RNN learns to predict the next character in a sequence (doesn’t necessary feel fluid and dynamic as a chatbot.)
  • single response outputs
  • cannot respond to a specific question (not smart enough. ouch.)
  • cannot generate response from scratch (picks from a fixed set.)

Learnings from what worked

  • responses are respective to their inputs (not completely nonsensical)
  • intuitive to use for creative writing
  • can be fun for long form texting

What could be improved

  • more sensitive, dynamic and emotionally intuitive response
  • more freedom to make exploratory conversations side-by-side with more emotional ones



NYU 2018–2020