Schedule

Monday, July 23rd

• Morning session:
  • 9:00: Registration and coffee/continental breakfast (5th floor)
  • 9:30: Hackathon overview and planning (Room 526)
• Afternoon session:
  • Hacking
  • 3:00: Coffee/tea break (5th floor)
  • 5:00: Daily sync-up meeting (Room 526)

Tuesday, July 24th

• Morning session:
  • 9:00: Coffee/continental breakfast (5th floor)
  • Reading and discussion group
• Afternoon session:
  • 3:00: Coffee/tea break (5th floor)
  • Hacking, shared task working group
  • 5:00: Daily sync-up meeting (Room 526)

Wednesday, July 25th

• Morning session:
  • 9:00: Coffee/continental breakfast (5th floor)
  • Hacking
• Afternoon session:
  • 3:00: Coffee/tea break (5th floor)
  • Hacking, shared task working group
  • 5:00: Daily sync-up meeting (Room 526)
• Evening: Optional Social Event at Navy Pier

Thursday, July 26th

• Morning session:
  • 9:00: Coffee/continental breakfast (5th floor)
  • Hacking and discussion
• Afternoon session:
  • 2:00: Elizabeth Clark, U Washington, "Collaborative Story Generation: Modeling and Design Considerations for Machine-in-the-Loop Creative Writing Systems"
  • 2:30: Kevin Gimpel, TTIC, "From Paraphrase Modeling to Controlled Generation"
  • 3:00: Coffee break
  • 3:30: Sam Wiseman, Harvard, "Learning Controllable and Interpretable Generation Models with Neural Templates"
  • 4:00-5:00: Shared task working group discussion/feedback
• Evening: Optional Social Event at Grant/Millennium Parks

Friday, July 27th

• Morning session:
  • 8:45: Mark Riedl, Georgia Tech, "Toward Interactive Improvisational Narrative Generation in Open Worlds"
  • 9:15: Melissa Roemmele, SDL, "The Role of Unpredictability in Automated Assistance for Creative Writing"
  • 9:45: Sameer Singh, UC Irvine, "On Knowledge Graph Embeddings, with Application to Generation"
  • 10:15: Coffee break
  • 10:45: Keynote: Andrew S. Gordon, University of Southern California: "Generating Creative Narratives About a Baseball Player, a Train, and a Heart"
• Lunch catered at TTIC (11:45am-1:15pm)
• Afternoon session:
  • 1:15: Larry Birnbaum, Northwestern, "Narrative Coherence and Semantic Structure"
  • 1:45: Kris Hammond, Northwestern, "The Difference between Words, Text, Language and Stories"
  • 2:15: Mohit Iyyer, UMass Amherst, "Generating QA Dialogs from Documents"
  • 2:45: Coffee break
  • 3:15: Hackathon team presentations
  • 4:15: Shared task presentation/discussion
  • 4:55: Break
  • 5:10: Jamie Brew, Botnik Studios, "Slow Markov songwriting: A Keynote concert"
  • 6:00: Closing remarks

Selected Talk Abstracts

• Keynote: Andrew S. Gordon, University of Southern California: "Generating Creative Narratives About a Baseball Player, a Train, and a Heart"

Story Creation Games, such as Rory's Story Cubes and the Tell Tale card game, require players to invent creative and coherent narratives from a set of unconnected elements assembled by random chance, e.g., the throw of a die or the draw of a card. Often producing comical and entertaining storylines, these games also demonstrate the remarkable human capacity for sense-making, where one's knowledge and experience is used to explain the co-occurrence of novel combinations of observations. In this talk, I describe our recent efforts to build a computer program that could successfully play story creation games. We view this task as an interpretation problem, where the aim is to identify a coherent narrative where each narrative element plays a structural role. Our approach is to solve this interpretation problem using logical abduction, searching for sets of narrative assumptions that logically entail each of the given narrative elements. The search proceeds by backchaining from narrative elements through a knowledge base of narrative and causal axioms expressed as first-order definite clauses, unifying assumptions wherever possible. After finding connected solutions that entail the given set of narrative elements, the structure of the proof graphs are then used to generate the natural language text representation of the interpretation. In this talk, I demonstrate this approach in generating eight creative narratives given the same set of three Tell Tale cards, depicting a train, a baseball player, and the symbol of a heart. These examples demonstrate that logical abduction is well-suited to this task, but also underscore the enormous knowledge bottleneck that must be overcome to play this game with arbitrary cards. I contrast our approach with recent efforts to generate narrative text using deep neural networks trained with narrative corpora, and discuss whether these approaches fundamentally change the nature of this knowledge bottleneck.

• Mark Riedl, Georgia Tech, "Toward Interactive Improvisational Narrative Generation in Open Worlds"

Improvisational storytelling involves one or more people interacting in real-time to create a story without advanced notice of topic or theme. Human improvisation occurs in an open-world that can be in any state and characters can perform any behaviors expressible through natural language. We explore the grand challenge of computational improvisational storytelling in open-world domains. The goal is to develop an intelligent agent that can sensibly co-create a story with one or more humans through natural language. We lay out some of the research challenges and explore a number of possible approaches.

• Melissa Roemmele, SDL, "The Role of Unpredictability in Automated Assistance for Creative Writing"

Based on the idea that human creativity involves unpredictable juxtapositions of ideas, some theoreticians have proposed that AI-based creativity support systems can utilize random algorithms to simulate this unpredictability. Several current text generation systems incorporate some degree of randomness in the generation process, such as machine learning models that produce words by randomly sampling from probability distributions learned from corpora. This unpredictability is undesirable for many types of generation tasks, but for systems that collaborate with people in authoring open-ended creative text, it can yield novel utterances that are perceived as creative. Quite obviously, however, there’s little benefit provided by a system that generates language randomly; the text must appear meaningful to the writer in order to have any chance of inspiring creativity. My goal is to start a discussion about the appropriate role of unpredictability in automated assistance for creative writing. I’ll provide some examples of existing applications that leverage this concept, and present some qualitative findings from my own work. Finally, I’ll pose the question of how to investigate this issue experimentally. Through this discussion I’ll emphasize that generation for creativity assistance should be examined differently from other generation tasks, and models should be designed based on the objectives of the writers that use them.