Abstract:One of the key challenges of automatic story generation is how to generate a long narrative that can maintain fluency, relevance, and coherence. Despite recent progress, current story generation systems still face the challenge of how to effectively capture contextual and event features, which has a profound impact on a model's generation performance. To address these challenges, we present EtriCA, a novel neural generation model, which improves the relevance and coherence of the generated stories through residually mapping context features to event sequences with a cross-attention mechanism. Such a feature capturing mechanism allows our model to better exploit the logical relatedness between events when generating stories. Extensive experiments based on both automatic and human evaluations show that our model significantly outperforms state-of-the-art baselines, demonstrating the effectiveness of our model in leveraging context and event features.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the key challenge in automatic story generation, that is, how to generate a long - length narrative that can maintain fluency, relevance and coherence. Despite the recent progress, current story - generation systems still have difficulty in effectively capturing context and event features, which has a profound impact on the generation performance of the model.
Specifically, the paper points out:
1. **Deficiencies of existing methods**:
- When generating stories, existing story - generation systems can generate seemingly reasonable single sentences, but there will be logical inconsistencies in the overall story. For example, if a car needs to be "repaired and replaced", then it should not be able to "continue driving"; "Ken" should not "drive at high speed" in the "snow"; if "Ken" "gets stuck in a ditch" or "loses traction", he should not be able to "drive a long distance".
- These problems stem from the failure to fully capture context features when tracking event sequences, because planned events usually lack background information (such as characters and scenes), and training stories may contain the same events but different reference stories, which may lead to confusion in the reasoning process.
2. **The proposed new method**:
- To solve these problems, the authors propose EtriCA (Event - Triggered Context - Aware Story Generation Augmented by Cross Attention), a new neural - generation model. EtriCA maps context features to event sequences through residual mapping and uses a cross - attention mechanism to improve the relevance and coherence of the generated story.
- This model allows for better utilization of the logical associations between events, thereby more effectively capturing context and event features when generating stories.
3. **Main contributions**:
- Propose a new task: generate a story according to the given guiding context and the planned event sequence.
- Introduce two datasets, in which multi - sentence narratives are paired with event sequences, for evaluating the coherence and relevance of the generated story.
- Propose a new neural - generation model EtriCA, which utilizes context and event sequences through an enhanced cross - attention mechanism and sentence - level representation learning.
- Through extensive experiments, demonstrate the superior performance of the proposed method in terms of fluency, coherence and relevance.
In summary, this paper aims to solve the deficiencies of existing story - generation systems in capturing context and event features by introducing the EtriCA model, thereby generating more fluent, relevant and coherent stories.