An investigation into NLP techniques for generating intelligent narrative feedback to support IDN authoring

Abstract

Authoring Interactive Digital Narratives (IDN) is challenging since past a certain size, it becomes hard to keep track of the user’s experience along all the different storylines. Natural Language Processing (NLP) provides us with the opportunity to generate such intelligent feedback that can help authors keep better track of the story space. This is what this PhD addresses.

In the first phase a systematic review of IDN literature is performed and list of User experience (UX) dimensions that could form the basis of feedback to authors is compiled. The second phase then maps these onto related areas of NLP research to see where these could be estimated automatically. This reveals 47 dimensions of UX covering 8 categories—23 of these map to 12 areas of NLP research, leading on to 5 specific examples of how they might help IDN authors: plotting emotional arcs, visualising emotion type and intensity, revealing the predictability of events, debugging internal story logic, and branch-wise summarization.

One of these NLP areas (Automatic Text Summarisation) is chosen for deeper investigation in Phase 3. A dataset is generated by simulating playthroughs of eight episodes from two narrative games - Before the Storm and Wolf Among Us using fan-created transcripts online. Annotations for extractive summarisation were created automatically by aligning extracts with fan-made abstractive summaries available online.

The dataset is released as open source for future researchers to train and test their approaches for IDN text. On applying common baseline extractive text summarization approaches to this dataset, several shortcomings in standard approaches are revealed when applied to narrative and interactive narrative datasets.

The last phase of this work experiments with using rationale-based learning with word-level and sentence-level rationales indicating the proximity of words and sentences to choice points. The results indicate that rationale-based learning can improve the ability of attention-based text summarisation models to create higher quality summaries that encode key narrative information better suggesting a promising new direction for narrative-based text summarisation models. In this way, this thesis takes a step toward generating authoring feedback to assist IDN authors as well as understanding the complexities and unique challenges posed by the domain.

More information

Identifiers

ORCID for Ashwathy T Revi: orcid.org/0000-0002-9936-8141

ORCID for David Millard: orcid.org/0000-0002-7512-2710

ORCID for Stuart Middleton: orcid.org/0000-0001-8305-8176

Catalogue record

Export record