Generating Different Story Tellings from Semantic Representations of Narrative

Elena Rishes, Stephanie M. Lukin, David K. Elson, Marilyn A. Walker

TL;DR

This paper introduces a pipeline for automatically generating multiple versions of the same story by bridging two specialized tools: Scheherazade, which encodes the semantic "meaning" of a story, and PERSONAGE, an NLG engine that can generate text in various "voices" or personalities. By mapping deep semantic structures to syntactic ones, the researchers demonstrate that we can move beyond rigid, hard-coded narratives toward interactive systems that adapt their storytelling style to different audiences.

What Problem Does This Paper Address?

In the "wild," storytelling is a social, dynamic act. Humans don't just recite facts; they adjust their tone, level of detail, and vocabulary based on who is listening. However, most interactive storytelling systems are brittle. They usually rely on hand-crafted text or simple "story trees" that make it difficult to change the narrative's "voice" without rewriting everything from scratch.

The technical challenge lies in the gap between content planning (what happens in the story) and surface realization (how it is actually said). Specifically, there has been a lack of automated methods to link rich, symbolic semantic representations of a narrative to the complex syntactic structures required by high-end Natural Language Generation (NLG) engines.

Methods

The authors developed an automated "bridge" between two existing frameworks:

  1. Scheherazade: An annotation tool that represents stories as a Story Intention Graph (SIG). This graph encodes events as predicate-argument structures (e.g., jump(fox)) anchored in ontologies like WordNet and VerbNet.
  2. PERSONAGE: An NLG engine that takes Deep Syntactic Structures (DSyntS) as input and applies stylistic parameters (based on Big Five personality traits) to produce varied outputs.

The Pipeline: * Extraction: The system uses the Scheherazade API to extract actions and timelines from the SIG encoding. * Mapping: It maps these semantic "frames" into lexico-syntactic classes (nouns, verbs, adjectives). * Transformation: It automatically builds a DSyntS dependency tree from the semantic data. * Realization: This tree is fed into PERSONAGE, which uses the RealPro surface realizer to turn the syntax into actual sentences.

By adjusting PERSONAGE’s 67 different parameters, the system can generate the same story in a "Shy" voice (using hedges like "it seems to me") or a "Laid-back" voice (using expletives or tag questions like "you see?").

Datasets

The study utilized DramaBank, a collection of story encodings.

  • Training/Development: One fable, "The Fox and the Grapes," was used to develop the translation rules.
  • Testing: The rules were then tested on the remaining 35 Aesop Fables in the collection (e.g., "The Lion and the Boar").

Key Insights and Ideas

  • Semantic Grounding is Key: Because Scheherazade’s SIG is already anchored in WordNet and VerbNet, it provides a linguistically "clean" foundation that makes it possible to map meaning to syntax across different stories without manual re-coding for every new plot.
  • One-to-Many Mapping: The core "aha moment" is that a single symbolic representation of a story can serve as the source for infinite stylistic variations. You don't need a new "plan" for a shy narrator versus a confident one; you just need a different set of realization rules.
  • The "Correctness" Baseline: Before trying to be "stylistic," the authors first had to prove the system could be "correct." They used Scheherazade’s own built-in (but rigid) realizer as a gold standard to ensure their automated pipeline wasn't losing the plot's meaning during the transformation.

Key Results

The researchers evaluated their system using BLEU scores (n-gram overlap) and Levenshtein Distance (edit distance) to compare PERSONAGE’s output against the baseline Scheherazade output.

  • Accuracy: On the test set of 35 fables, the system achieved a Levenshtein Mean of 72 and a BLEU score of 0.32. While lower than the development set (0.59 BLEU), the results were statistically similar to the baseline in terms of closeness to the original fables.
  • Qualitative Variety: The system successfully produced distinct versions of stories. For example, where a formal version says "The fox didn't obtain the grapes," a "Laid-back" version might say "He damn was hungry... the fox failed to get the group of grapes, okay?"

Limitations and Future Work

  • Linguistic Nuance: The current system struggles with irregular verb forms (generating "openned" or "wifes") and lacks sophisticated handling of articles (a vs. the) and complex tenses.
  • Evaluation Metrics: The authors acknowledge that BLEU and Levenshtein are "correctness" metrics, not "quality" metrics. A "Shy" retelling should be different from the original, so high similarity isn't always the goal. Future work will involve human subjective judgments.
  • Interactive Voices: The current study focused on the narrator's voice. The next step is applying these stylistic models to individual characters within the story to give them unique personalities during dialogue.

Summary updated

Abstract

In order to tell stories in different voices for different audiences, interactive story systems require: (1) a semantic representation of story structure, and (2) the ability to automatically generate story and dialogue from this semantic representation using some form of Natural Language Generation (nlg). However, there has been limited research on methods for linking story structures to narrative descriptions of scenes and story events. In this paper we present an automatic method for converting from Scheherazade’s story intention graph, a semantic representation, to the input required by the personage nlg engine. Using 36 Aesop Fables distributed in DramaBank, a collection of story encodings, we train translation rules on one story and then test these rules by generating text for the remaining 35. The results are measured in terms of the string similarity metrics Levenshtein Distance and BLEU score. The results show that we can generate the 35 stories with correct content: the test set stories on average are close to the output of the Scheherazade realizer, which was customized to this semantic representation. We provide some examples of story variations generated by personage. In future work, we will experiment with measuring the quality of the same stories generated in different voices, and with techniques for making storytelling interactive.
BibTeX
@inproceedings{rishes2013generating,
  author = {Elena Rishes and Stephanie M. Lukin and David K. Elson and Marilyn A. Walker},
  title = {{Generating Different Story Tellings from Semantic Representations of Narrative}},
  booktitle = {ICIDS},
  year = {2013},
  url = {https://openreview.net/forum?id=rn8zCelvLWe},
}

Loading related content...

Figure