Story Realization: Expanding Plot Events into Sentences

Prithviraj Ammanabrolu, Ethan Tien, Wesley Cheung, Zhaochen Luo, William Ma, Lara J. Martin, Mark O. Riedl

Abstract: Neural network based approaches to automated story plot generation attempt to learn how to generate novel plots from a corpus of natural language plot summaries. Prior work has shown that a semantic abstraction of sentences called events improves neural plot generation and allows one to decompose the problem into: (1) the generation of a sequence of events (event-to-event) and (2) the transformation of these events into natural language sentences (event-to-sentence). However, typical neural language generation approaches to event-to-sentence can ignore the event details and produce grammatically-correct but semantically-unrelated sentences. We present an ensemble-based model that generates natural language guided by events. We provide results—including a human subjects study—for a full end-to-end automated story generation system showing that our method generates more coherent and plausible stories than baseline approaches.

Introduction

Automated story generation is a complex task often hindered by the tendency of neural models to "hallucinate" or ignore plot constraints. To manage this complexity, researchers typically split the process into two distinct stages: Event-to-Event planning, which generates a high-level plot sequence, and Event-to-Sentence realization, which translates these abstract events into readable natural language.

Events are represented as semantic abstractions—specifically 5-tuples consisting of a subject, verb, direct object, preposition, and modifier. While this abstraction reduces data sparsity, standard neural translation models (like Seq2Seq) often fail at the realization step. They frequently act as simple language models, generating generic text that ignores the input event's specific constraints. This paper introduces an ensemble-based "story realization" system designed to bridge this gap.

Methodology: The Story Realization Pipeline

The proposed system utilizes a cascading ensemble of five models, balancing the trade-off between faithfulness to the plot event and the creativity of the prose. The pipeline processes sentences through an "eventification" stage, utilizes working and long-term memory for entity consistency, and employs a slot-filler to re-insert specific names and entities into generated templates.

The story realization pipeline, illustrating the transformation of input sentences into abstract events and the subsequent generation of new sentences via memory and slot-filling modules.

The Ensemble Components

  1. Retrieve-and-Edit (RetEdit): This model searches training data for similar events, retrieves the corresponding sentence, and edits it to fit the new event, preserving complex structures found in human-written stories.
  2. Sentence Templating: A rule-based approach using a simplified grammar and an AWD-LSTM language model to fill blanks. This is the most faithful model but can be formulaic.
  3. Monte-Carlo Beam Search (MCBS): A Seq2Seq model using a custom scoring function to reward the inclusion of keywords from the input event.
  4. Finite State Machine (FSM) Constrained Beams: This forces the inclusion of specific event tokens by guiding the beam search through a finite state machine.
  5. Vanilla Seq2Seq: A standard baseline used as a fallback when other models fail to meet confidence thresholds.

Experimental Setup

The system was trained on a specialized corpus of Sci-Fi TV Plot Summaries scraped from Wikia, containing 2,276 stories from series such as Doctor Who and The X-Files. Sentences were preprocessed into generalized 5-tuples using VerbNet and WordNet to reduce vocabulary size and highlight action patterns.

Results and Discussion

Quantitative evaluation showed that the "Full Ensemble" outperformed all individual models and the Seq2Seq baseline, achieving the lowest Perplexity (70.179) and the highest BLEU-4 (0.0481) and ROUGE-4 scores.

Qualitative results from a human subjects study on Amazon Mechanical Turk further validated the approach. Participants rated the ensemble-generated stories significantly higher than the baseline across several key metrics, including grammar, genre consistency, and overall enjoyability.

Average human study scores from Amazon Mechanical Turk comparing the baseline model to the proposed ensemble across multiple qualitative metrics.

An analysis of model utilization revealed that while RetEdit was preferred for known data (~95%), the system relied more heavily on Monte-Carlo Beam Search (~41%) when processing novel, AI-generated plots. This suggests that neural "creativity" is essential for handling the unpredictability of automated plot planners.

Conclusion

By treating story realization as a guided "denoising" problem, this ensemble approach ensures that generated text remains faithful to the intended plot without sacrificing grammatical fluidity. Future work involves incorporating more sophisticated language models and improving long-range narrative consistency to ensure characters do not perform contradictory actions across the story arc.

Summary updated

Abstract

Neural network based approaches to automated story plot generation attempt to learn how to generate novel plots from a corpus of natural language plot summaries. Prior work has shown that a semantic abstraction of sentences called events improves neural plot generation and and allows one to decompose the problem into: (1) the generation of a sequence of events (event-to-event) and (2) the transformation of these events into natural language sentences (event-to-sentence). However, typical neural language generation approaches to event-to-sentence can ignore the event details and produce grammatically-correct but semantically-unrelated sentences. We present an ensemble-based model that generates natural language guided by events. We provide results—including a human subjects study—for a full end-to-end automated story generation system showing that our method generates more coherent and plausible stories than baseline approaches 1.
BibTeX
@inproceedings{ammanabrolu2020story,
  author = {Prithviraj Ammanabrolu and Ethan Tien and Wesley Cheung and Zhaochen Luo and William Ma and Lara J. Martin and Mark O. Riedl},
  title = {{Story Realization: Expanding Plot Events into Sentences}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  url = {https://openreview.net/forum?id=cGF5-Ymz7ho},
}
Figure