Human-Guided Evolutionary-Based Linguistics Approach For Automatic Story Generation

Abstract

Existing automatic storytelling approaches are limited in terms of the levels of creativity, coherence and interestingness in their generated stories. We conjecture that the ability to stochastically evolve stories using evolutionary computation (EC) methods is a solution to these problems. EC relies on an implicit self-feedback loop in which stories generated in one iteration contribute to subsequent generations of improved stories and this process can rely on human feedback to guide evolutionary dynamics.

This thesis is to introduce a framework for a human-guided evolutionary-based linguistics approach for automatically generating coherent, novel and interesting stories using computers.

Two proposals are discussed in this thesis. The first relies on an evolutionary story generation approach using an existing `fabula' story structure with a parameterised version of tree adjoining grammar as the story formalism. The results demonstrate that the system is able to produce coherent and complex stories with multiple characters, branches and long-distance dependency. Moreover, the system can collect good stories by introducing a closed feedback loop, through human interaction, to the computational story generation process implemented through grammar-guided genetic programming. However, this proposal highlights some problems resulting from using a story structure that holds only a single level of complexity and the computational and unrealistic complexity associated with relying on full human-in-the-loop story evaluation.

The second proposal attempts to find solutions to these two problems. We extract a hierarchical dependence network from an existing story, represent it formally using a special data structure and then propose a permutation-based chromosome to encode a generated story from it. When compared with the original existing story, it possesses different degrees of flashback, participant arrangement and event order. Subsequently, multi-objective hybrid EC is applied to collect good stories. The problem of the pure human-in-the-loop evaluation in the first proposal is alleviated by a semi-interactive evaluation method in which an approximated human story evaluation model automatically evaluates the generated stories. Human-based story evaluation experiments are carried out to collect subjective data to calculate the approximated human evaluation model and verify the multi-objective evolutionary storytelling approach. The model successfully guides the evolution to collect stories with improved quality.