Story Generation Based on Multi-granularity Constraints.

CICAI (3)(2022)

引用 0|浏览0
Generating a coherent story is a challenging task in natural language processing, which requires maintaining the coherence of plots and inter-sentence semantics throughout the generated text. While existing generation models can generate texts with good intra-sentence coherence, it is still difficult to plan a coherent plot and inter-sentence semantics throughout the text. In this paper, we propose a novel multi-granularity constraints text generation model, which constrains at the token level and sentence level, respectively. For the plot incoherence issue, the token-level constraint is added, which is a new plot guidance method to maintain the coherence of the plot while avoiding the introduction of extra exposure bias. For the problem of semantic incoherence, an auxiliary task of modelling the semantic relations between sentences is designed on the sentence-level constraint. Extensive experiments have shown that our model can generate more coherent stories than the baselines.
Story generation, Multi-granularity constraints, Pre-training model
AI 理解论文