Guided Neural Language Generation for Automated Storytelling

Prithviraj Ammanabrolu, Ethan Tien, Wesley Cheung, Zhaochen Luo, William Ma, Lara J. Martin, Mark O. RiedlACL 2019

TL;DR

This paper tackles the challenge of generating coherent and engaging stories from abstract plot points (events). They propose an ensemble of neural models for the "event-to-sentence" task, outperforming previous sequence-to-sequence approaches by better preserving event semantics and generating more natural language. The work also demonstrates a full end-to-end story generation pipeline, integrating their event-to-sentence model with an existing event-to-event generator.

What Problem Does This Paper Address?

Automated story generation is a complex task that involves creating a sequence of plot points that are consistent, maintain long-term dependencies, and leverage common-sense knowledge. While early approaches relied on manual knowledge engineering, recent work has shifted towards neural networks trained on large story corpora. A key challenge in neural story generation is the sparsity of data, where unique sentences are common, making it difficult for models to learn robust mappings. To address this, researchers have proposed abstracting plot points into "events" (tuples of subject, verb, object, etc.). This decomposes the problem into two stages: generating a sequence of events (event-to-event) and converting these events into natural language sentences (event-to-sentence). However, existing event-to-sentence models, particularly vanilla sequence-to-sequence networks, often ignore the input event, producing grammatically correct but semantically disconnected sentences. This paper aims to improve the event-to-sentence conversion by developing a more robust guided language generation system that preserves the semantic meaning of events while producing engaging narratives.

Methods

The core of this paper is the development of an ensemble-based model for the event-to-sentence task. Recognizing that no single model perfectly balances semantic fidelity and sentence quality, the authors combine four specialized models with a baseline sequence-to-sequence model as a fallback. The four models are:

  • Retrieve-and-Edit (RetEdit): This model first retrieves a similar event-sentence pair from the training data and then "edits" it to generate a new sentence based on the input event. It excels at retaining complex sentence structures but can struggle with events distant from the training data.
  • Sentence Templating: This approach uses a simplified grammar derived from event structures and a Bidirectional LSTM language model to generate sentences. It guarantees that all event tokens are present in the output, ensuring semantic preservation, but can lead to structurally similar sentences.
  • Monte Carlo Beam Search: An adaptation of beam search that introduces an additional scoring term based on "playouts" (sampling words from the vocabulary). This method generates more diverse and grammatically correct sentences but doesn't guarantee retention of all input event tokens.
  • Finite State Machine Constrained Beams: This model uses finite state machines to guide beam search, enforcing the presence of input event tokens in the generated sentence. It aims to balance flexibility with adherence to lexical constraints, though it can sometimes fail to produce a sentence within a given length or lose proper grammar.

These models are combined in a cascading sequence, with confidence scores determining which model is queried. The ensemble is designed to leverage the strengths of each individual model and mitigate their weaknesses. The paper also implements an event-to-event model (a policy gradient deep reinforcement learner) to generate event sequences for the full pipeline experiment.

Datasets

The primary dataset used for training and evaluation is a collection of science fiction TV show plot summaries scraped from the fandom wiki service wikia.com. This corpus contains 2,276 stories, with an average length of 89.23 sentences, spanning 11 shows. The data was pre-processed to simplify alien names. Sentences were split and "eventified" (converted into event tuples) following the methodology of Martin et al. (2018). The dataset was split into training, validation, and testing sets in an 8:1:1 ratio.

Key Insights and Ideas

  • The "event-to-sentence" problem is a hard balancing act: No single neural model can simultaneously guarantee perfect semantic preservation from the event and generate highly engaging, natural language sentences.
  • Ensembling is key: Combining specialized models, each optimized for different aspects of the task (semantics vs. fluency vs. diversity), is a powerful strategy to achieve a better overall outcome.
  • Vanilla sequence-to-sequence models are insufficient: They often act as simple language models, ignoring the crucial semantic guidance provided by events, leading to disconnected narratives.
  • Event representation is crucial but not a silver bullet: While events help decompose the problem and reduce data sparsity, the conversion to natural language remains a significant challenge.
  • The cascade approach offers computational efficiency: By querying models in a specific order and stopping when a confidence threshold is met, the ensemble avoids unnecessary computation.

Key Results

  • The ensemble model outperforms individual event-to-sentence models in terms of perplexity and BLEU-4 scores, demonstrating its effectiveness in balancing semantic meaning and sentence quality.
  • The Retrieve-and-Edit model was heavily utilized on the test set, likely due to similar data distributions between training and testing.
  • The Sentence Templating model produced the shortest sentences, while the Finite State Machine model generated longer ones due to its constraints.
  • The paper successfully demonstrates a full end-to-end story generation pipeline, showing how their event-to-sentence ensemble can be integrated with an event-to-event generator to produce complete stories from unseen input data.

Limitations and Future Work

  • Limitations:
    • The Retrieve-and-Edit model's performance is heavily dependent on the availability of similar examples in the training data, which is limited by data sparsity.
  • The Finite State Machine Constrained Beams model, while aiming for constraints, can sometimes fail to produce a sentence within a fixed time horizon or may lose proper grammar.
  • The ensemble's confidence thresholds were tuned on the validation set, and their optimal values might vary for different datasets or domains.
  • Future Work:
    • The authors suggest incorporating other state-of-the-art sentence-filling techniques into the ensemble to further address model weaknesses.
  • The work serves as a stepping stone for research in sentence expansion or more advanced event-to-sentence tasks.
  • Exploring different event representations or more sophisticated ways to combine model outputs could lead to further improvements.

Summary updated

BibTeX
@inproceedings{ammanabrolu2019guided,
  author = {Prithviraj Ammanabrolu and Ethan Tien and Wesley Cheung and Zhaochen Luo and William Ma and Lara J. Martin and Mark O. Riedl},
  title = {{Guided Neural Language Generation for Automated Storytelling}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure