NEWS: We are currently seeking a PhD student to work on this project, initially focusing on the two topics listed below. Applicants should have a solid background in computer science and an interest in natural language technology. If you are interested, please contact Patrik or Hanna.
The plot (or narrative) of a story has some similarity with a plan, as it is usually defined in classical AI planning. This has led researchers to investigate planning-based approaches to automatic story generation (see, for example, the work of Porteous et al. and their demo). However, the planning approach requires as input a formal description of the "world" in which the story is to take place - it's characters, places, objects, their relations, and the things they can do. Creating such a world model is a great burden on system designers, and limits the creativity that story generation systems can exhibit. On the other hand, machine learning approaches can produce generative models from data in a mostly unsupervised manner, but these models so far are only able to replicate the surface appearance of fiction and fail to generate a coherent plot (sunspring is an excellent example in point).
The ultimate goal of automated story generation is to release unbounded creativity while still making sense. To approach this problem, we look to methods for learning structured semantic knowledge from abundantly available sources, such as text, as well as already existing collections of such knowledge that is on the web. Initiatives such as conceptnet and read the web offer large databases of facts and relations (both abstract and concrete) obtained from text mining and crowdsourcing, while ontologies can provide knowledge about specific areas, or general things like time (see, for example, vocab.org). However, a type of sematic knowledge that is crucial to narratives and that is currently not easily available is knowledge about events (such as what causes them to happen, what their consequences are, and how they chain together).
Realising the vision of unlimited narrative creativity is a great task, which cannot be completed within a single student project. Below are some current specific topics within the scope of this project. One of these can form the basis for an individual research, honours or masters project, while a PhD will span multiple topics and go beyond these starting points.
Topic: Time line extraction
A first step towards understanding the plot of a text is to identify the events that it describes and order them into a consistent time line. Extracting ordered event sequences is also a prerequisite for applying some event model learning methods.
Students interested in this topic should read the honours thesis by William James.
Topic: Event model learning
Event models capture knowledge about possible events, such as their participants, prerequisites, consequences, and relations to other events, and are a key ingredient in a planning approach to narrative generation. There are many different approaches to learning event or action models from data, including from examples of event sequences or by mining text collections.
Students working on this project need to have a background in computer science and good programming skills.
- Follow the links!
- Haslum, "Narrative Planning: Compilations to Classical Planning", Journal of AI Research, vol. 44, p. 383-395, 2012, also available on-line from JAIR.
- Julie Porteous, Fred Charles and Marc Cavazza, "NetworkING: using Character Relationships for Interactive Narrative Generation", International Conference on Autonomous Agents and Multiagent Systems, 2013 (available here).
Previous students' work on this topic:
- David Cowley carried out a user study in interactive fiction and player modelling (honours thesis, 2015).
- Emily Rodrigo implemented a Markov chain model for text generation, and explored the question if coupling it with a topic model could generate text that was more coherent and sensible (report, 2017).
- William James (co-supervised with Hanna Suominen) implemented the beginnings of an NLP pipeline for extracting action/event models from text (honours thesis, 2017).
- Musha Wen investigated the use of the conceptnet database, combined with NLP extraction techniques, to find event relations that may be used to create story planning operators (report, 2018).
- Louis Carlin implemented and evaluated another technique for extracting event models from text (report, 2019).
- Jiaqi Zhang (report, 2019) and Debashish Chakraborty (thesis, 2019) both studied the shortcomings of GAN-based text generation.