Neural story generation models have two significant challenges: (1) coherence over narrative structure, especially long-range dependencies, and (2) emotional coherence and consistency, generally producing redundant or incoherent narration. A new, emotionally intelligent two-stage short story generation model is presented by combining GPT-2 with a tailored FNET model, a light transformer architecture substituting standard self-attention with Fourier Transform layers to improve semantic and emotional relationship capture in text. The first stage employs GPT-2 to generate a list of input candidate sentences, a question, an answer, and an emotional state. The candidate sentences are then filtered using an emotion classifier from DistilRoBERTa to keep only those that adhere to a desired emotional tone. The filtered sentences are then fed into a fine-tuned FNET model, which examines inter-sentence relationships and enforces emotional coherence to generate a coherent and emotionally engaging narrative. An empirical comparison using three benchmark datasets demonstrates the system's superiority over earlier state-of-the-art approaches. The FNET model achieves 0.3093 in BLEU-1, outperforming Plan-and-Write (0.0953) and T-CVAE (0.2574), with an enhanced narrative quality and lexical coherence with human-written narratives. The story coherence and emotion retention accuracies are 85%, 67%, and 60% for Visual7W, ROCStories, and Cornell Movie Dialogs datasets.
Kachare et al. (Sat,) studied this question.