Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Event Representations for Automated Story Generation with Deep Neural Nets (1706.01331v3)

Published 5 Jun 2017 in cs.CL, cs.AI, cs.LG, and cs.NE

Abstract: Automated story generation is the problem of automatically selecting a sequence of events, actions, or words that can be told as a story. We seek to develop a system that can generate stories by learning everything it needs to know from textual story corpora. To date, recurrent neural networks that learn LLMs at character, word, or sentence levels have had little success generating coherent stories. We explore the question of event representations that provide a mid-level of abstraction between words and sentences in order to retain the semantic information of the original data while minimizing event sparsity. We present a technique for preprocessing textual story data into event sequences. We then present a technique for automated story generation whereby we decompose the problem into the generation of successive events (event2event) and the generation of natural language sentences from events (event2sentence). We give empirical results comparing different event representations and their effects on event successor generation and the translation of events to natural language.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (7)
  1. Lara J. Martin (14 papers)
  2. Prithviraj Ammanabrolu (39 papers)
  3. Xinyu Wang (186 papers)
  4. William Hancock (2 papers)
  5. Shruti Singh (13 papers)
  6. Brent Harrison (30 papers)
  7. Mark O. Riedl (57 papers)
Citations (221)
View on Semantic Scholar

Summary

An In-Depth Analysis of Event Representations for Automated Story Generation with Deep Neural Nets

The paper "Event Representations for Automated Story Generation with Deep Neural Nets" by Martin et al. addresses the problem of open story generation by leveraging deep learning techniques to learn story structures from text. The authors emphasize the limitations of prior symbolic planning approaches for story generation, such as their dependency on human-knowledge engineering and restriction to predetermined domains. The paper posits that neural networks, specifically a recurrent encoder-decoder framework, can better handle the open story generation problem by learning from a comprehensive corpus without the need for preconfigured domain models.

Event Representation

At the core of the paper is the notion of an event representation, which captures the semantic essence of a story while minimizing event sparsity. The authors propose using a 4-tuple event representation ⟨s,v,o,m⟩\langle s, v, o, m \rangle, where vv denotes the verb, ss the subject, oo the object, and mm a modifier. This abstraction seeks to encapsulate key story elements while enhancing potential overlap across different stories. By reducing the uniqueness of events and increasing the number of observations of each event type, the research aims to foster the predictive capabilities of neural networks, essential for the generation of coherent stories.

Results and Evaluation

The paper presents a series of empirical results comparing different event representations using a recurrent neural network model. The authors find that generalizing events improves prediction capabilities, as indicated by lower perplexity scores compared to original sentence or word-level representations. Additionally, the results exhibit the benefit of maintaining a context or history (event bigrams) in generating more plausible story continuations.

The paper shows that the abstraction level in event representation directly impacts the predictive and generative performance of the neural network. Furthermore, incorporating auxiliary information, such as genre data, can refine these predictions. While the BLEU scores remain relatively low, they underline the inherent complexity and challenge in translating event sequences into meaningful narratives, contrasted by improvements in perplexity.

Challenges in Event-to-Sentence Mapping

The transition from events to human-readable sentences is identified as a fundamental challenge. Abstraction improves story generation but transforms the narrative into non-intuitive event sequences that must be converted back to natural language. The authors introduce an "event2sentence" network to restore human-readability by training on pairs of events and the sentences they are derived from. The paper observes that generalization aids in event generation, but simultaneously complicates the reverse translation due to the loss of sentence-specific details.

Implications and Future Directions

The implications of this research reverberate through both practical and theoretical domains. Practically, advancements in story generation could revolutionize content creation industries, educational tools, and interactive storytelling platforms. Theoretically, this work raises questions around the trade-offs between abstraction and specificity in natural language processing tasks, opening avenues for further exploration into memory-augmented neural networks and reinforcement learning paradigms for guiding story progression.

The authors also propose future enhancements, including the development of memory mechanisms to manage entity consistency and content variability across longer narratives, to improve story coherence and detail fidelity.

In conclusion, "Event Representations for Automated Story Generation with Deep Neural Nets" represents a substantial step towards open story generation, underpinned by a novel approach to event representation. It highlights the nuanced balance required between abstraction and detail in automated narrative construction and sets the stage for the integration of more advanced AI storytelling models leveraging deeper semantic and pragmatic understanding.