In the realm of story generation, there exists a dichotomy between instruction-tuned large language models (LLMs) and older symbolic approaches such as planning. While LLMs excel in generating stories in response to open-ended user requests, they often lack diversity and struggle with long-term story coherence. On the other hand, symbolic approaches can produce more diverse plot outlines but are limited by fixed character action templates. To bridge the gap between these two approaches, researchers have explored combining their strengths while mitigating their weaknesses. One innovative approach involves using answer set programming (ASP) to provide a higher-level and abstract symbolic specification of high-level story structure. By leveraging ASP to guide and diversify LLM-based story generation, researchers have demonstrated the ability to produce more diverse stories compared to unguided LLMs. The proposed method involves generating a wide range of potential story outlines through an answer set program encoding various storytelling goals. Subsequently, a random outline is selected from this set and expanded into a complete textual story using an LLM, based on a provided high-level story premise. This novel approach not only showcases the compactness and flexibility of ASP-based outline generation but also highlights the diversity benefits achieved through semantic similarity analysis of generated stories. By combining the narrative capabilities of LLMs with the structural advantages of symbolic outline generation, this research opens new avenues for enhancing story generation processes and improving overall storytelling quality.
Introduction:
Storytelling has been an integral part of human culture since the beginning of time. From oral traditions to written literature, stories have served as a means of entertainment, education, and communication. With the advancement of technology, artificial intelligence (AI) has entered the realm of storytelling through story generation. This involves using computer programs to automatically create narratives in response to user input or predefined prompts.
In recent years, large language models (LLMs) such as GPT-3 have gained popularity for their ability to generate coherent and realistic text based on open-ended user requests. However, these models often lack diversity and struggle with long-term coherence when it comes to generating complex narratives. On the other hand, older symbolic approaches like planning can produce more diverse plot outlines but are limited by fixed character action templates.
To bridge this gap between LLMs and symbolic approaches, researchers have explored combining their strengths while mitigating their weaknesses. One innovative approach involves using answer set programming (ASP) to provide a higher-level and abstract symbolic specification of high-level story structure.
The Dichotomy Between Instruction-Tuned LLMs and Symbolic Approaches:
Instruction-tuned LLMs excel at generating stories in response to open-ended user requests due to their vast knowledge base acquired from training on large datasets. These models use natural language processing techniques to understand user input and generate relevant text that follows grammatical rules.
However, when it comes to creating longer narratives with multiple characters and plot twists, LLMs often struggle with maintaining coherence and producing diverse outcomes. This is because they rely heavily on statistical patterns learned from training data rather than understanding underlying concepts or themes.
On the other hand, symbolic approaches involve explicitly defining rules or constraints for generating narrative structures based on predefined templates or scripts. While this allows for more control over the overall story structure and potential for diversity in plot outlines, it also limits creativity as these templates are fixed.
Combining LLMs and Symbolic Approaches through ASP:
To overcome the limitations of both LLMs and symbolic approaches, researchers have explored using ASP as a bridge between the two. ASP is a declarative programming language that allows for high-level specification of complex problems by defining rules and constraints.
In this context, an answer set program encodes various storytelling goals such as character development, plot progression, and conflict resolution. These goals are then used to generate a wide range of potential story outlines.
The generated outlines are then fed into an LLM-based story generation system, which expands them into complete textual stories based on a provided high-level story premise. This approach not only leverages the narrative capabilities of LLMs but also incorporates the structural advantages of symbolic outline generation.
Benefits of Using ASP in Story Generation:
One major benefit of using ASP in story generation is its compactness and flexibility. By encoding storytelling goals in a declarative manner, it allows for easy modification or addition of new goals without affecting the overall structure.
Moreover, by generating multiple potential outlines through ASP, there is increased diversity in the narratives produced by LLMs compared to unguided models. This is achieved through semantic similarity analysis where similar or redundant outcomes are filtered out to ensure diverse outputs.
Implications for Future Research:
The use of ASP in combination with LLMs has opened up new avenues for enhancing story generation processes and improving overall storytelling quality. However, there is still room for improvement and further research in this area.
One potential direction could be exploring more sophisticated methods for selecting or ranking generated outlines based on their relevance to the given premise or user input. Additionally, incorporating user feedback during the outline selection process could lead to more personalized and engaging narratives.
Conclusion:
In conclusion, while there exists a dichotomy between instruction-tuned large language models (LLMs) and older symbolic approaches such as planning when it comes to story generation, researchers have found ways to bridge this gap and combine their strengths while mitigating their weaknesses. The use of answer set programming (ASP) as a higher-level and abstract symbolic specification for story structure has shown promising results in terms of diversity and coherence in LLM-based story generation. This innovative approach not only showcases the compactness and flexibility of ASP but also highlights the benefits achieved through semantic similarity analysis. With further research, this method could potentially revolutionize the field of AI storytelling and lead to more dynamic and engaging narratives.
