TL;DR
This document comprises the proceedings of the Second Workshop on Storytelling at ACL 2019, showcasing research at the intersection of human storytelling and artificial intelligence. The papers explore various aspects of story generation, understanding, and analysis, highlighting the multidisciplinary nature of this field and its growing importance for grounding AI in human experience.
What Problem Does This Paper Address?
The workshop addresses the fundamental role of storytelling in human cognition and communication, and how this can be leveraged and replicated within artificial intelligence. While storytelling is a core human activity, computationally generating, understanding, and analyzing narratives presents significant challenges. This workshop brings together researchers from diverse fields to explore these challenges, aiming to advance AI's ability to engage with and produce human-like narratives. The papers investigate aspects such as story structure, components, expression, and the application of AI techniques to these areas.
Methods
The papers presented at this workshop employ a wide range of methods, reflecting the multidisciplinary nature of the field. These include:
- Neural and Linguistic Approaches: Many papers utilize deep learning models, recurrent neural networks (RNNs), and transformer architectures for tasks like story generation and understanding. Linguistic theories and methods are also integrated to capture narrative structure and meaning.
- Grounded Story Generation: Some research focuses on generating stories that are grounded in specific personas or real-world knowledge, moving beyond purely abstract text generation.
- Understanding Narrative Structures: Techniques are explored to analyze and understand the underlying schemas, plot structures, and emotional arcs within narratives, including those found in social media and fan fiction.
- Multimodal Storytelling: Approaches are investigated for composing stories that involve multiple modalities, such as text and images.
- Evaluation Metrics: The workshop implicitly addresses the challenge of evaluating generated stories, with papers proposing or utilizing various metrics to assess coherence, creativity, and other qualitative aspects.
Datasets
The datasets used in the papers are diverse and reflect the varied applications of storytelling research:
- General Text Corpora: Standard large text datasets are likely used for pre-training language models.
- Narrative Texts: Collections of novels, short stories, and literary works are employed for analyzing narrative structures and generating fictional content.
- Social Media Data: Platforms like Twitter are used to study narrative generation and understanding in real-time, informal contexts.
- Fan Fiction: Datasets derived from fan fiction communities are utilized to study emotional communication and narrative trends.
- Picture Books: Datasets of picture books are used for research on multimodal story generation.
- Movie Plot Summaries: These are used for predicting movie success.
- NaNoGenMo: This dataset, likely from the "National Novel Generation Month" event, is used for studying narrative generation methods.
- Custom Datasets: Some papers may introduce or curate specific datasets tailored to their particular research questions, such as those for persona-based generation or character-centric storytelling.
Key Insights and Ideas
- Storytelling as a Bridge for AI: The workshop strongly emphasizes that storytelling is not just an artistic endeavor but a crucial component for developing more human-like and understandable AI systems. By grounding AI in the way humans naturally communicate and make sense of the world through stories, we can build more robust and relatable AI.
- The Multidisciplinary Nature of Storytelling: The breadth of research presented highlights that effective storytelling in AI requires collaboration across linguistics, computer science, cognitive science, and even the humanities. No single discipline holds all the answers.
- Beyond Coherence: The Importance of Persona and Emotion: Generating coherent text is a baseline. The workshop shows a clear trend towards generating stories with specific personas, emotional depth, and groundedness in real-world contexts, which are key elements of compelling human narratives.
- The Challenge of Evaluation: Quantifying the "goodness" of a story remains a significant challenge. While metrics for fluency and coherence exist, capturing creativity, emotional impact, and narrative arc is an ongoing area of research.
- Leveraging Diverse Data Sources: From formal literature to informal social media and fan fiction, researchers are exploring how to extract and generate narratives from a wide array of sources, each presenting unique challenges and opportunities.
Key Results
The proceedings showcase a variety of results across different research areas:
- Story Generation: Papers demonstrate progress in generating stories that are more coherent, contextually relevant, and sometimes exhibit specific stylistic or persona-driven characteristics. For example, some models can compose picture books or generate more engaging story endings.
- Story Understanding: Research shows advancements in identifying narrative structures, classifying genres, detecting everyday scenarios, and recognizing personality traits within literary texts.
- Emotional Analysis: Studies explore how to analyze and potentially generate stories with specific emotional content, particularly within the context of fan fiction.
- Predictive Modeling: One paper demonstrates the potential of deep learning models to predict movie success based on plot summaries.
- Quantitative Improvements: While specific numbers vary by paper, the collective work suggests incremental but significant improvements in metrics related to story generation quality and narrative understanding tasks.
Limitations and Future Work
- Evaluation Metrics: A recurring theme is the difficulty in objectively evaluating the quality of generated stories. Future work will likely focus on developing more nuanced and human-aligned evaluation metrics that capture aspects like creativity, emotional resonance, and narrative impact.
- Data Scarcity and Bias: For specialized storytelling tasks (e.g., persona-based generation, specific genres), the availability of large, high-quality datasets can be a limitation. Future research may involve creating new benchmarks or developing techniques that are less data-hungry.
- Long-Term Coherence and Planning: While progress has been made in generating coherent short narratives, maintaining long-term plot consistency and complex character development over extended stories remains a significant challenge.
- Ethical Considerations: As AI becomes more adept at generating persuasive narratives, ethical considerations regarding misinformation, manipulation, and the impact of AI-generated content will become increasingly important.
- Multimodality: Integrating different modalities (text, image, audio) for richer storytelling experiences is an active area for future exploration.
- Grounding and Common Sense: Further work is needed to imbue AI storytellers with common sense reasoning and a deeper understanding of the world, enabling them to generate more believable and relatable narratives.
Summary updated