[Paper Review] Inferring the Reader: Guiding Automated Story Generation with Commonsense Reasoning

Xiangyu Peng, Siyan Li+2 more|arXiv (Cornell University)|May 4, 2021

Topic Modeling19 references4 citations

TL;DR

This paper proposes Commonsense-inference Augmented neural StoryTelling (CAST), a framework that enhances story coherence by integrating commonsense reasoning into neural story generation. CAST uses COMET2020 to infer character goals, emotions, and intentions from generated sentences, filtering only those continuations that align with inferred reader mental models—resulting in significantly more coherent, on-topic, fluent, and enjoyable stories than strong baselines in both single- and two-character settings.

ABSTRACT

Transformer-based language model approaches to automated story generation currently provide state-of-the-art results. However, they still suffer from plot incoherence when generating narratives over time, and critically lack basic commonsense reasoning. Furthermore, existing methods generally focus only on single-character stories, or fail to track characters at all. To improve the coherence of generated narratives and to expand the scope of character-centric narrative generation, we introduce Commonsense-inference Augmented neural StoryTelling (CAST), a framework for introducing commonsense reasoning into the generation process with the option to model the interaction between multiple characters. We find that our CAST method produces significantly more coherent, on-topic, enjoyable and fluent stories than existing models in both the single-character and two-character settings in three storytelling domains.

Motivation & Objective

• To address the lack of commonsense reasoning and plot coherence in neural story generation models.
• To enable coherent, character-centric storytelling across multiple characters by modeling goal hierarchies and character intentions.
• To bridge the gap between neural language model generation and human reader comprehension by simulating reader mental models.
• To improve narrative fluency and enjoyment by enforcing logical, goal-driven story progression.
• To develop a scalable framework that supports both single- and multi-character story generation with minimal architectural changes.

Proposed method

• The system starts with a prompt and a specified main character, then generates candidate story continuations using a pre-trained language model.
• For each candidate, COMET2020 infers four commonsense dimensions: character goals, emotions, intentions, and potential outcomes.
• A candidate is accepted only if its inferred attributes match those from the previous sentence with sufficient consistency (using a threshold-based matching mechanism).
• The accepted continuation becomes the new prompt, and the process repeats until a story of desired length is generated.
• The framework supports both single-character and two-character narratives via interleaved generation and dynamic character switching.
• The method relies on a Markovian assumption, comparing only the current event to the immediately preceding one, though it maintains global coherence effectively.

Experimental results

Research questions

• RQ1 Can commonsense reasoning improve the coherence and fluency of neural story generation beyond likelihood-based generation?
• RQ2 How does modeling character goals and intentions affect story quality in multi-character narratives?
• RQ3 Can a filtering mechanism based on reader mental model inference outperform baseline methods in story generation?
• RQ4 To what extent does the system maintain topic relevance and narrative consistency over long sequences?
• RQ5 How does the integration of COMET2020 inferences improve story enjoyment and logical progression?

Key findings

• CAST significantly outperforms strong baselines such as C2PO and Goldfarb-Tarrant et al. (2020) in story coherence, fluency, and topic relevance across three storytelling domains.
• The system produces more enjoyable and logically consistent stories by enforcing alignment with inferred reader mental models through commonsense inferences.
• In both single- and two-character story generation, CAST achieves higher scores on human evaluation metrics, including coherence, on-topicness, and fluency.
• The framework demonstrates robustness in maintaining global narrative coherence despite a local, Markovian comparison strategy.
• CAST generates more diverse stories than C2PO due to its use of COMET2020 only for filtering, not for sentence generation.
• The performance of CAST is directly tied to the quality of COMET2020’s commonsense inferences, suggesting future gains with improved inference models.