























































































AH. Kachare et al. /Future Technology                                                                           November 2025| Volume 04 | Issue 04 | Pages 
138-145 

138 
 

 

 

Article 

FNet-GPT: Fourier-based lightweight transformer 
for emotion-aware text generation using GPT 
Atul Haribhau Kachare1,2*, Chandrashekhar Goswami1, Ashutosh Gupta1, D.S. Chouhan3  
1Department of Computer Science & Engineering, Sir Padampat Singhania University, Udaipur, Rajasthan, India 
2Department of Computer Engineering, Shah & Anchor Kutchhi Engineering College, Mumbai, Maharashtra, India 
3Department of Mathematics, Sir Padampat Singhania University, Udaipur, Rajasthan, India 

A R T I C L E   I N F O 
 

Article history: 
Received 15 June 2025  
Received in revised form 
25 July 2025 
Accepted 13 August 2025 
 
Keywords:  
Story generation, Emotionally aware narration, 
FNET, GPT-2, Natural Language Processing 
 
*Corresponding author 
Email address: 
atul.kachare@spsu.ac.in 
 
 
DOI: 10.55670/fpll.futech.4.4.12 

A B S T R A C T 
 

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. 

1. Introduction 
Text generation is a process of automatically creating 

text that sounds like it was written by a human, which is 
trending in the world of artificial intelligence (AI). It has its 
roots in the early days of computational linguistics research 
[1]. Nowadays, with modern text generation models and deep 
learning methods like transformers [2], we can write a text 
that's not only coherent but also contextual. It employs 
various applications, including writing and creating content 
for power chatbots, as well as generating code [3]. However, 
to move ahead, we must consider the ethical considerations 
[4]. Short story generation, a part of text generation, is about 
crafting tight narratives with gripping plots and characters. 
With early models like the n-gram language [5], creating 
stories with sense was difficult. Nevertheless, with deep 
learning, especially with the introduction of models like GPT-
3 [6], machines can tell creative stories, complete with rich 
characters and surprising twists. Still, we have many issues, 
such as making sure stories are coherent, diverse, and used 
ethically [7]. Text generation covers a broad spectrum of 
creating different types of text sentences, paragraphs, or code 
while ensuring grammar and coherence. It can be used for 

machine translation, summarization, or question answering. 
Story generation, a subset of text generation, focuses 
explicitly on crafting narratives. It requires linguistic fluency 
and understanding of plot structures, character development, 
and thematic elements. Story generation aims to create 
engaging and imaginative tales that evoke emotions and 
captivate readers. This paper is categorized as follows: 
Section 1 introduces the research area. The background of the 
problem and a discussion of different approaches are given in 
Section 2, whereas Section 3 outlines a literature review. 
Section 4 details the proposed system and its methodology. 
Section 5 outlines the experimental setup, including datasets, 
baselines, and metrics, and presents the results obtained. The 
final section concludes the article by discussing the findings 
and potential future directions. References to relevant works 
are provided at the end. 

2. Background 
The historical trajectory of story generation is deeply 

rooted in human culture, evolving from oral traditions and 
written literature to computational approaches [8]. Early 
computer programs utilized rule-based systems, followed by 

 

 

Future Technology 

Open Access Journal 

https://doi.org/10.55670/fpll.futech.4.4.12 

November 2025| Volume 04 | Issue 04 | Pages 138-145 

Journal homepage: https://fupubco.com/futech 

 
ISSN 2832-0379 

mailto:atul.kachare@spsu.ac.in
https://doi.org/10.55670/fpll.futech.4.4.12
https://fupubco.com/futech


AH. Kachare et al. /Future Technology                                                                           November 2025| Volume 04 | Issue 04 | Pages 138-145 

139 
 

more sophisticated methods like case-based reasoning and 
planning, exemplified by MINSTREL [9] and BRUTUS [10]. 
The rise of machine learning, particularly RNNs, LSTMs [11], 
and transformer-based models like GPT [12] and BERT [13], 
has significantly advanced the field. The current landscape is 
dominated by large language models (LLMs) like GPT-3 [3] 
and interactive storytelling. So, the different template-based 
systems [14], rule-based approaches [15], case-based 
reasoning [9], planning models [16], and even simulation 
techniques [17] were explored. With the arrival of neural 
language models like GPT-3, the story generation process is 
transformed into creative and diverse narratives. However, 
there are still some hurdles in controlling what these models 
generate and ensuring they are ethical. Looking ahead, we see 
a blend of strategies to craft stories that grab attention and 
ensure we keep those ethical considerations in mind. 

3. Literature survey 
The research conducted by Fan A. and Lewis M. [18] 

aimed to enhance fluency and coherence by utilizing a 
hierarchical model with self-attention, but encountered 
issues such as tokenization and text repetition. Xu J. and Ren 
X. [19] focused on making sentences connected semantically 
for better coherence, using a Seq2Seq model, but struggled 
because there were not enough human-annotated examples 
in the real-world datasets they were working with. Yao L. and 
Peng N. [20] introduced a hierarchical framework with 
explicit storyline planning, but the model struggled with off-
topic content and inconsistencies. Wang T. and Wan X. [21] 
developed a model for generating coherent plots within 
incomplete stories using a T-CVAE, but faced limitations in 
generating story endings. Chen G. and Liu Y. [22] proposed 
generating an outline to bridge the gap between title and 
story, resulting in more extensive narratives, but capturing 
cross-sentence dependencies remained a challenge. Zhang Y. 
and Shi X. [23] utilized a Knowledge Graph to generate the 
image captioning, but acknowledged limitations due to the 
offline construction of the graph by computing cosine 
similarity. Chen G. and Liu Y. [24] create the outline from the 
training data with a title and story using a variational neural 
network, but they generate only a single-sentence outline, 
which restricts its ability to support complex or long-form 
story structures. Brahman F. and Chaturvedi S. [25] focused 
on generating emotionally aware stories, but the lack of large-
scale annotated story corpora posed a challenge. Tan B. and 
Yang Z. [26] proposed a progressive generation method for 
long text passages, but the need to expand vocabulary while 
maintaining relevance and accuracy remained a limitation. 
Min K. and Dang M. [27] proposed generating short stories 
from images using RNNs and an encoder-decoder model, but 
with grammar and emotional expression limitations. Wu C. 
and Wang J. [28] introduced the ICPGN model for generating 
classical Chinese poems from images, but its reliance on a 
specific dataset limits it. Liu Y. and Huang Q. [29] proposed 
the SSAP model for generating story endings based on context 
and sentiment using ChatGPT-2. Jin Y. and Kadam V. [30] 
presented SCRATCHPLOT, a method for generating stories 
using pre-trained language models without fine-tuning, but it 
requires significant post-processing. Chen Y. and Li R. [31] 
proposed a co-creative visual storytelling method that allows 
user control over events and emotions, but needs better 
prompt formats and emotion classifiers. Khan L. and Gupta V. 
[32] focused on generating coherent stories using keywords 
and genre-based inputs by optimizing the Hugging Face GPT-
2 model, but the influence of the title or keyword on the 

narrative remains limited. Based on the literature survey, 
some research gaps are identified as follows: 
• Failure to manage the proper emotional context: Most 

models fail to guarantee that the output text is always tied 
to a target emotion throughout the story. 

• Inadequate modelling of long-range dependencies: Current 
systems do not have strong mechanisms for modelling and 
preserving coherence across sentences or pieces of a story. 

• Recurrent behaviours and narrative flow inconsistencies 
are commonly found in generative models because of 
shallow attention mechanisms or the absence of semantic 
feedback loops. 

• Lack of a cohesive system that identifies emotions and 
increases coherence: Few research efforts have attempted 
to combine emotional filtering with semantic structuring in 
a single, effective system. 

These identified gaps highlight the need for a unified story 
generation system to simultaneously manage emotional 
control and narrative coherence using efficient, scalable 
techniques. Although deep learning models like GPT-2 have 
revolutionized natural language generation, they remain 
limited in their ability to generate stories that are both 
emotionally consistent and narratively coherent. Current 
systems often produce emotionally neutral content and 
exhibit weak cross-sentence dependencies, leading to 
repetitive or disconnected storylines. Moreover, the lack of 
integration between emotion classification and sequence 
modeling further restricts the expressive quality of generated 
narratives. These limitations create a gap in developing AI-
generated stories that genuinely resonate with human 
emotions and maintain a logical narrative flow. Bridging this 
gap requires a hybrid approach that fuses emotional filtering 
with semantically aware text generation. To address these 
limitations, this paper proposes a two-stage hybrid 
architecture that combines the generative power of GPT-2, 
the emotion recognition capability of DistilRoBERTa, and the 
semantic modeling efficiency of FNET. The key objectives of 
this research are: 
• To ensure emotional alignment by filtering GPT-2-

generated sentences using DistilRoBERTa. 
• To enhance narrative coherence by modeling global 

dependencies through a Fourier-based FNET architecture. 
• To empirically validate the system on benchmark datasets 

(Cornell Movie Dialogs, Visual7W, and ROCStories). 
• To evaluate performance using BLEU and METEOR metrics 

and emotion retention accuracy. 

4. Proposed system 
Having performed an extensive review of existing 

literature, we identify two long-standing issues in the domain 
of automatic narrative generation: 
• Repetition and Inconsistency: The majority of modern 

systems lack consistent character behavior and consistent 
story progression, instead generating repetitive or 
contradictory content. 

• Lack of Cross-Sentence Dependency Modeling: The current 
models are limited by their failure to capture long-distance 
semantic and affective relations between sentences, 
impacting narrative coherence. 

We have developed a hybrid model, as shown in Figure 1, that 
combines GPT-2, DistilRoBERTa, and a specially crafted FNET 
model to tackle specific challenges. FNET is a Transformer 
encoder that mixes via the Fourier transforms instead of 
using the usual self-attention method. It helps effectively 
capture those global dependencies. 



AH. Kachare et al. /Future Technology                                                                           November 2025| Volume 04 | Issue 04 | Pages 138-145 

140 
 

 
Figure 1. Proposed system for short story generation 

The reason to use FNET: 
• Global Context Awareness: Many attention mechanisms 

become complex as the input grows. Nevertheless, FNET's 
Fourier transform encoder is more scalable. It can handle 
complete sequences. 

• Improved Consistency: Another great thing about FNET is 
that it keeps the narrative on track. It minimizes those 
annoying moments where the story drifts or characters 
seem inconsistent. It does this by holding onto those long-
range dependencies. 

• Reduced Computational Expense: FNET skips the whole 
attention mechanism altogether. It operates with a time 
complexity of O(n log n), which is significantly better than 
the O(n²) we see in standard Transformers. The system, as 
proposed, works in two phases.  

A comparative analysis of FNet with Linformer, Longformer, 
and Performer is presented in Table 1, which outlines the 
trade-offs across computational efficiency, memory footprint, 
and real-world performance. It supports the design decision 
to employ FNet for coherent and emotionally grounded 
narrative generation. FNet bypasses self-attention by 
applying a two-dimensional Discrete Fourier Transform 
(DFT) across token and embedding dimensions, enabling fast 
global mixing. Linformer, while highly efficient with O(n) 
complexity, uses low-rank projections that may weaken its 
ability to retain subtle narrative transitions or emotional 
nuances required in story generation. Longformer is 
optimized for document-level tasks using a sliding window 
and sparse global attention, which can be less effective for 
capturing inter-sentence dependencies in shorter narratives. 
Performer approximates self-attention using kernel-based 
methods and random projections, offering scalability but with 
increased computational cost. FNet thus provides a 
compelling balance: faster than self-attention, semantically 
expressive, and well-suited for maintaining long-range 
dependencies, which are crucial for generating coherent and 
emotionally aligned narratives. 

 

 

 

 

 

 

 

 

 

 

4.1 Stage 1: Emotionally aware sentence generation 
• The system is presented with a question, a corresponding 

answer, and a target emotional state (e.g., happiness, fear, 
anger). 

• Sentence Generation: A pre-trained GPT-2 model produces 
some potential sentences based on the inputs provided. 

• Emotion Filtering: Every sentence is passed through a 
DistilRoBERTa-based emotion classifier. Sentences 
belonging to the target emotion only are kept. 

• Selection Mechanism: A beam search algorithm produces 
diverse sentence forms with maximum semantic and 
emotional appropriateness. 

We employed the Emotion English DistilRoBERTa-base 
model [33], classifying text into seven categories: anger, 
disgust, fear, joy, neutral, sadness, and surprise. It is built by 
fine-tuning DistilRoBERTa-base on a balanced subset (~20k 
samples) drawn from six diverse English-language datasets, 
including GoEmotions, Crowdflower, ISEAR, MELD, etc. Each 
emotion category is represented by approximately 2,811 
examples, with 80% used for training and 20% for evaluation. 
On held-out data, the model attains an accuracy of 66%. 

4.2 Stage 2: FNET-based coherent story generation 
The FNet encoder replaces the standard self-attention 

mechanism of transformers with a 2D Discrete Fourier 
Transform (DFT2), achieving efficient global mixing with 
reduced computational complexity. This section formally 
describes the mathematical operations that govern the 
encoding process. Given a sequence of tokens 𝑥𝑥 =
 [𝑥𝑥1, 𝑥𝑥2, … , 𝑥𝑥𝑛𝑛]. Each token is embedded in a vector, 𝐸𝐸𝑖𝑖 =
𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸(𝑥𝑥𝑖𝑖) ∈ 𝑅𝑅𝑑𝑑. So the input matrix becomes: 

𝑋𝑋 ∈  𝑅𝑅𝑛𝑛×𝑑𝑑                              (1) 

Where n: sequence length, d: embedding dimension, each row 
𝑥𝑥𝑖𝑖  correspond to the embedding of the ith token. 
To preserve the order of tokens in the input sequence, we add 
fixed sinusoidal positional encodings 𝑃𝑃 ∈  𝑅𝑅𝑛𝑛×𝑑𝑑and type 
encodings 𝑇𝑇 ∈  𝑅𝑅𝑛𝑛×𝑑𝑑  

𝑋𝑋′ = 𝑋𝑋 + 𝑃𝑃 + 𝑇𝑇                                                            (2) 

These encodings allow the model to distinguish between 
tokens at different positions without learning position-
specific parameters. Instead of using quadratic-complexity 
self-attention, FNet applies a 2D Discrete Fourier Transform 
(DFT2) across both the token and embedding dimensions of 
the input: 

𝑍𝑍 = 𝑅𝑅𝐸𝐸 �𝐹𝐹𝐹𝐹𝑇𝑇2(𝑋𝑋′)�𝑅𝑅𝑛𝑛×𝑑𝑑                                              (3) 

 

 

 

 

 

 

 

 

 

 

 

Table 1. Comparative analysis of efficient transformer variants across time complexity, memory usage, parallelizability, and practical 
applicability 

Model Time Complexity Memory Usage Parallelizability Real-World Use 

FNET O(n log n) Low High High-speed on GPUs; minimal memory 

Lin 
former O(n) Low Moderate Needs pre-defined projection; sensitive to 

rank 

Long 
former 

O(n) (locally), O(n²) for 
global tokens 

Medium-High Medium Scales well on long docs, less for short 
sequences 

Per 
former O(n) Medium High Random projections increase training 

overhead 
 

 



AH. Kachare et al. /Future Technology                                                                           November 2025| Volume 04 | Issue 04 | Pages 138-145 

141 
 

Here, FFT2(.) is a 2D Fourier Transform applied to the matrix 
(1st DFT across Rows: Sequence dimension and then 2nd DFT 
across columns: Feature dimension), and Re(.) extracts the 
real part to keep the result compatible with downstream 
layers. This operation transforms the input into the frequency 
domain, enabling global interaction between all tokens via 
frequency components. The core idea is that Fourier mixing 
captures long-range dependencies through global frequency 
patterns without explicitly computing pairwise attention 
scores. It replaces the standard attention computation with 
an efficient and non-learned operation that reduces the time 
complexity from 𝑂𝑂(𝑛𝑛2) to 𝑂𝑂(𝑛𝑛 log 𝑛𝑛), as demonstrated in the 
original FNet work [34]. At this stage, we combine the output 
of the Fourier Transform block Z with the original embedding 
input 𝑋𝑋′. It helps retain the original input signal and makes 
the model more stable during training. After this addition, a 
Layer Normalization operation is applied to the result. 
LayerNorm standardizes the summed output along the 
feature dimensions to improve convergence and avoid 
internal covariate shift. 

𝑌𝑌 = 𝐿𝐿𝐿𝐿𝐿𝐿𝐸𝐸𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐸𝐸 (𝑍𝑍 + 𝑋𝑋′)                                              (4) 

The normalized output is passed through a standard position-
wise feedforward network: 

𝐻𝐻 = 𝑅𝑅𝐸𝐸𝐿𝐿𝑅𝑅(𝑌𝑌𝑌𝑌1 + 𝐸𝐸1)𝑌𝑌2 + 𝐸𝐸2                                                   (5) 

Where, 𝑌𝑌1,𝑌𝑌2  ∈  𝑅𝑅𝑛𝑛×𝑑𝑑 are learnable weight matrices and 
𝐸𝐸1, 𝐸𝐸2  ∈  𝑅𝑅𝑑𝑑 are biases. ReLU adds non-linearity to improve 
representational power. 
Finally, another residual connection is added between the 
input to the Feedforward block and the output of the 
Feedforward block. It helps the model combine local 
(feedforward) and global (Fourier) features. After adding 
them, LayerNorm is applied again to stabilize and standardize 
the representation. 

𝐻𝐻′ = 𝐿𝐿𝐿𝐿𝐿𝐿𝐸𝐸𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐸𝐸 (𝑌𝑌 + 𝐻𝐻)                                    (6) 

Once the final representation 𝐻𝐻′ is ready and mapped to the 
vocabulary space using a linear output projection. It is done 
by multiplying it with a learnable weight matrix: 

𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 = 𝐻𝐻′𝑌𝑌0 + 𝐸𝐸0                                     (7) 

The model applies the softmax function to the logits to 
generate the next token and get probabilities. Then, it uses 
argmax to choose the most likely token: 

𝐿𝐿�𝑡𝑡 = arg max�𝑆𝑆𝐿𝐿𝑆𝑆𝐿𝐿𝐸𝐸𝐿𝐿𝑥𝑥(𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿)�                         (8) 

The final output is a short story composed of emotionally 
consistent and semantically connected sentences generated 
from the filtered GPT-2 outputs and refined through FNET's 
encoding-decoding mechanism. 

This system leverages FNET to pick up on the subtle 
meanings and feelings, along with GPT-2's amazing 
storytelling abilities. Such a multi-step journey; craft short 
stories that resonate with emotion while responding to 
specific questions and answers. The first stage is all about 
creating sentences that are aware of emotions. It begins with 
taking in a question, an answer, and an emotional context. 
GPT-2 works to weave a narrative that feels relevant and 
coherent. Then, the emotion classifier, which has been trained 
on data tagged with various emotions, helps pinpoint the 
specific feelings in the sentences that get generated. The beam 
search algorithm explores all kinds of word combinations, 
figuring out which arrangements could form sensible 
sentences. After generating all those options, the system goes 

through them again with the emotion classifier to ensure that 
the final picks match the intended emotional tone. In the 
latter part of the system, we have used an advanced FNET 
model as shown in Figure 2, in which a typical self-attention 
layer is replaced by a Fourier Transform layer [35] to run 
smoother and quicker. Self-attention is a mechanism that 
helps the model understand how different words relate to 
each other in a sentence [36]. The Fourier Transform method 
breaks down a signal into frequency components, helping us 
identify long-range correlations in the input sequence, which 
can be beneficial. The model's structure has two key parts to 
focus on: the encoder and the decoder. The encoder takes in 
the input sequence and, at the end of its job, comes up with a 
context vector. Then, the decoder takes over, grabs that 
context vector, and turns it into the output sequence. 
Positional embeddings are added to the input and output 
sequences to help us understand where each word belongs. 

 
 

Figure 2. FNET model 

5. Evaluation of the proposed system 
We evaluate the system to see how well it could write an 

emotionally engaging short story. We examine three different 
benchmark datasets, each one having some unique 
challenges. The system was evaluated against some 
established baseline models to measure the quality and 
coherence of the stories. In the following sections, we will 
delve into the setup of our experiments, the results we 
obtained, and the insights gained from those findings. 

5.1 Dataset 
So, we decided to work with three benchmark datasets 

for our experiments to see how well the proposed system 
performs. First up, we have the Cornell Movie-Dialogs Corpus 
[37]. It is a pretty massive collection, with all these fictional 
conversations pulled from movie scripts, totaling around 
220,579 exchanges between 10,292 pairs of characters. Next 
is the Visual7W dataset [38], which comes from COCO images 
and is packed with 327,939 question-answer pairs. Plus, it 
features 1,311,756 human-generated answer choices and 
61,459 object groundings. While Visual7W is originally a 
visual QA dataset, in our framework, it plays a crucial role in 
enabling visually grounded story generation. We adapt the 
dataset by using its image-question-answer triples to train 
our Visual Question Answering and Visual Question 



AH. Kachare et al. /Future Technology                                                                           November 2025| Volume 04 | Issue 04 | Pages 138-145 

142 
 

Generation (parts of a larger system), producing prompts and 
emotions that are used to guide the narrative generation 
process. It allows the system to generate stories that are not 
only coherent but also contextually aligned with visual 
content. Furthermore, the emotional relevance of each image 
is inferred via Visual Sentiment Analysis, helping to tailor the 
tone and affective quality of the generated narrative. Lastly, 
we looked at the ROCStories dataset [39], a resource for NLP 
research, with 98,162 five-sentence stories about everyday 
life. It has been designed to help models understand and 
generate stories by capturing causal and temporal 
relationships. 

5.2 Baseline models 
We compare our models with the following: 1) T-CVAE 

[21] leverages transformers and a variational autoencoder to 
learn story patterns and generate new, coherent stories, with 
the ability to incorporate additional input for guidance. 2) 
Plan&Write [20] operates in two stages: planning a storyline 
and generating the story text based on that plan, offering 
better control over story structure and coherence. The 
comparison aims to showcase the effectiveness and 
advancements of the proposed approach in story generation. 

5.3 Metrics 
We have chosen to utilize the BLEU (Bilingual Evaluation 

Understudy) [40] and METEOR [40] metrics to evaluate the 
quality of the generated stories. BLEU is one of the most 
popular machine translation metrics that determines the 
similarity between a human reference translation and a 
machine translation. BLEU verifies the similarity between the 
n-grams (set of n words) in the target text and the reference 
text—the more similar n-grams, the greater the BLEU score, 
which shows greater content preservation. 
The BLEU score formula is derived based on precision and a 
penalty for brevity. The formula is: 

𝐵𝐵𝐿𝐿𝐸𝐸𝑅𝑅 = 𝐵𝐵𝑃𝑃 × exp �1
𝑛𝑛
∑ log𝑝𝑝𝑖𝑖𝑛𝑛
𝑖𝑖=1 �                                 (9) 

However, BLEU does not consider the stems and synonyms of 
words. To overcome these limitations, METEOR uses a 
weighted F1-score and penalty function. 

𝐹𝐹𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 =
10� 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚ℎ𝑒𝑒𝑒𝑒 𝑢𝑢𝑛𝑛𝑖𝑖𝑖𝑖𝑖𝑖𝑚𝑚𝑚𝑚

𝑢𝑢𝑛𝑛𝑖𝑖𝑖𝑖𝑖𝑖𝑚𝑚𝑚𝑚 𝑖𝑖𝑛𝑛 ℎ𝑦𝑦𝑦𝑦𝑦𝑦𝑚𝑚ℎ𝑒𝑒𝑖𝑖𝑒𝑒
�� 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚ℎ𝑒𝑒𝑒𝑒 𝑢𝑢𝑛𝑛𝑖𝑖𝑖𝑖𝑖𝑖𝑚𝑚𝑚𝑚

𝑢𝑢𝑛𝑛𝑖𝑖𝑖𝑖𝑖𝑖𝑚𝑚𝑚𝑚 𝑖𝑖𝑛𝑛𝑖𝑖𝑒𝑒𝑖𝑖𝑒𝑒𝑖𝑖𝑒𝑒𝑛𝑛𝑚𝑚𝑒𝑒
�

� 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚ℎ𝑒𝑒𝑒𝑒 𝑢𝑢𝑛𝑛𝑖𝑖𝑖𝑖𝑖𝑖𝑚𝑚𝑚𝑚
𝑢𝑢𝑛𝑛𝑖𝑖𝑖𝑖𝑖𝑖𝑚𝑚𝑚𝑚 𝑖𝑖𝑛𝑛𝑖𝑖𝑒𝑒𝑖𝑖𝑒𝑒𝑖𝑖𝑒𝑒𝑛𝑛𝑚𝑚𝑒𝑒

�+9� 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚ℎ𝑒𝑒𝑒𝑒 𝑢𝑢𝑛𝑛𝑖𝑖𝑖𝑖𝑖𝑖𝑚𝑚𝑚𝑚
𝑢𝑢𝑛𝑛𝑖𝑖𝑖𝑖𝑖𝑖𝑚𝑚𝑚𝑚 𝑖𝑖𝑛𝑛 ℎ𝑦𝑦𝑦𝑦𝑦𝑦𝑚𝑚ℎ𝑒𝑒𝑖𝑖𝑒𝑒

�
                     (10) 

 

𝑃𝑃𝐸𝐸𝑛𝑛𝐿𝐿𝑃𝑃𝐿𝐿𝐿𝐿 = 0.5 × �𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁 𝑜𝑜𝑜𝑜 𝑐𝑐ℎ𝑁𝑁𝑛𝑛𝑢𝑢𝑢𝑢 𝑖𝑖𝑛𝑛 𝐻𝐻𝐻𝐻𝐻𝐻𝑜𝑜𝑡𝑡ℎ𝑁𝑁𝑖𝑖𝑢𝑢
𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁 𝑜𝑜𝑜𝑜 𝑁𝑁𝑚𝑚𝑡𝑡𝑐𝑐ℎ𝑁𝑁𝑑𝑑 𝑁𝑁𝑛𝑛𝑖𝑖𝑢𝑢𝑁𝑁𝑚𝑚𝑁𝑁𝑢𝑢

�                 (11) 

 

𝑀𝑀𝐸𝐸𝑇𝑇𝐸𝐸𝑂𝑂𝑅𝑅 = 𝐹𝐹𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆  ×  (1 − 𝑃𝑃𝐸𝐸𝑛𝑛𝐿𝐿𝑃𝑃𝐿𝐿𝐿𝐿)                   (12) 

5.4 Experimental setup 
We utilize a pre-trained GPT-2 model to generate a wide 

range of sentences. It will take a question and its answer, then 
churn out several ways to say it. Now, the emotion classifier 
will look at all those candidate sentences and determine 
which hits the emotional mark for the story we are crafting. 
This way, our sentences will help shape the narrative's overall 
vibe. At the core of our story-generating system is the FNET 
model — a transformer architecture known for its high 
efficiency. We have an encoder with five layers that'll sift 
through the input sentences, getting a good grasp of their 
meaning and context. Then, a decoder with eight layers will 
spin all that info into a coherent narrative. We will also use 
positional embeddings to keep track of word order in both the 

input and the output. The FNET model will learn from stories, 
improving at creating narratives that make sense and 
resonate emotionally based on the input and the feelings we 
want to convey. Moreover, we use the Adam optimizer, a 
standard in deep learning, to keep everything on track during 
training. It will help fine-tune the model so that what it 
generates is as close to the stories we aim for. 

5.5 Result 
This section presents the experimental evaluation's 

outcomes, showcasing the proposed system's performance in 
generating emotionally aware short stories. The results are 
analyzed in comparison to baseline models, highlighting the 
strengths and limitations of the proposed approach. Table 2 
presents the results of training a machine learning model 
called FNet on three different datasets: 
• Cornell's Movie Dataset: This dataset likely contains movie 

dialogues or scripts. The FNET model achieved an accuracy 
of 0.60 (60%) on this dataset. It suggests that the model 
performs moderately well in understanding or generating 
movie dialogue. There might be room for improvement, as 
40% of the model's predictions were incorrect. 

• Visual7w Dataset: This dataset probably contains images 
paired with questions and answers about the visual content 
of the images. The FNET model achieved a high accuracy of 
0.85 (85%) on this dataset, indicating that it is quite 
effective at understanding visual content and answering 
questions about it. 

• ROC-Stories: This dataset likely contains short stories or 
narratives. FNET achieved an accuracy of 0.67 (67%) on 
this dataset. It suggests that the model performs 
reasonably well in understanding or generating short 
stories, although there is still potential for improvement. 
 

Table 2. Accuracy of trained FNet on different benchmark datasets 

Dataset Accuracy 

Cornell's Movie Dataset 0.60 

Visual7W Dataset 0.85 

ROC-Stories Dataset 0.67 

 

5.6 Discussion 
Table 3 compares three models, Plan-and-Write, T-CVAE, 

and the Proposed System, based on BLEU and METEOR 
metrics, which evaluate how closely the generated text 
matches the reference human-written text. 
BLEU score analysis 
• Plan-and-Write: This model has the lowest scores across all 

BLEU metrics. It suggests that its generated text has the 
least overlap in individual words, bigrams, and trigrams 
compared to the human-generated reference text. 

• T-Cave: It shines regarding BLEU-2 scores, which means it 
is pretty good at churning out pairs of words matching the 
reference text. However, its scores for BLEU-1. It hints that 
it might struggle with individual words. 

• The Proposed model: It grabs the top spot for BLEU-1, 
which means it nails those individual word matches well. 
Its BLEU-2 score is not the absolute highest, but they are 
still solid compared to other models. 

• Overall: So, T-Cave is excellent for generating those word 
pairs, but the Proposed model seems to strike a better balance 
across all the BLEU metrics, which suggests it might come 
closer to sounding like a human would write, especially 



AH. Kachare et al. /Future Technology                                                                           November 2025| Volume 04 | Issue 04 | Pages 138-145 

143 
 

regarding vocabulary choices and how the phrases fit 
together. 
METEOR score analysis 
• Plan-and-Write achieves a METEOR score of 0.266, 

indicating the lowest semantic alignment among the three. 
• T-CVAE slightly improves with a METEOR of 0.278, 

reflecting a modest gain in semantic closeness. 
• The Proposed System obtains the highest METEOR score of 

0.291, indicating better overall alignment with human-
written references, not just in exact word matches, but also 
in meaning and structure. 

Table 3. Comparative analysis of BLUE and METEOR scores for 
baseline and proposed system for ROCStories dataset 

Model 
BLEU Score 

METEOR 
BLEU-1 BLEU-2 

Plan-and-Write [20] 0.0953 0.0159 0.266 

T-CVAE [21] 0.2574 0.0987 0.278 

Proposed System 0.3093 0.0871 0.291 

 

Since the output generations from T-CVAE and Plan-and-
Write were not publicly available, we were unable to perform 
statistical significance testing. However, our proposed model 
consistently outperforms the baselines across multiple 
metrics and datasets. We acknowledge the importance of such 
testing and plan to include it in future work when comparable 
outputs are accessible. We conducted a small-scale human 
evaluation of our proposed model's ability to generate 
emotionally grounded short stories to supplement automatic 
metrics. Five participants rated five stories using a 5-point 
Likert scale. Results consistently showed good scores in 
emotional relevance and coherence, indicating that, by using 
simple language, the FNet-GPT model effectively produces 
coherent and emotionally resonant narratives. 
Comparison with other existing emotional story 
generation models: 

FNet-GPT differentiates itself from AffectStory [41], 
PlotMachines [42], and StoryGAN [43] by providing a unique 
blend of explicit emotional control, enhanced textual 
coherence, and superior computational efficiency for 
narrative generation. While AffectStory relies on theoretical 
cognitive models for emotion and PlotMachines focuses on 
plot adherence with implicit emotional outcomes, FNet-GPT 
integrates an emotion filtering mechanism and a Fourier 
Transform-based FNet for emotional consistency and 
efficient long-range dependency handling in the text. 
Furthermore, unlike StoryGAN, primarily a text-to-image 
visualization model, FNet-GPT strictly focuses on generating 
high-quality, emotionally nuanced textual stories. Compared 
to general-purpose controllable models like CTRL, FNet-GPT 
offers more specialized and direct emotional conditioning 
with a more efficient architectural design 𝑂𝑂(𝑛𝑛 log 𝑛𝑛) vs. 𝑂𝑂(𝑛𝑛2) 
complexity, making it particularly well-suited for emotion-
aware text generation. 

5.7 Error analysis 
Despite employing an emotion classifier 

(DistilRoBERTa) during the filtering stage, we found that a 
few generated stories from the Cornell dataset deviated from 
the intended emotional tone. Similarly, in all datasets, a few 
outputs showed weak narrative transitions, including abrupt 
shifts in events or character actions, indicating issues with 

coherence. Future system versions could benefit from fine-
tuning the emotion classifier on story-specific datasets to 
address these limitations and enhance filtering precision. 
Additionally, incorporating a story planning module could 
help maintain logical flow and improve narrative structure 
throughout the generated text. 

6. Conclusion  
This paper introduces a novel, emotionally intelligent short 
story generation architecture by combining the robust 
language generation ability of GPT-2 with the semantic 
richness and sparsity of a specially designed FNet model. The 
combined model addresses some of the most significant 
issues of existing literature, i.e., redundancy in a narrative, 
inconsistency in emotion, and weak inter-sentence 
dependency. An empirical investigation of three well-
established benchmark sets—Cornell Movie Dialogs, 
Visual7W, and ROCStories—unequivocally demonstrates the 
efficacy of the suggested method. In particular, the system 
achieves 85% accuracy on Visual7W, 67% on ROCStories, and 
60% on Cornell, manifestly demonstrating its ability to 
maintain emotional coherence and semantic consistency in 
diverse contexts. Another comparison of BLEU scores 
confirms that the suggested model outperforms existing 
state-of-the-art baselines, with a BLEU-1 score of 0.4093, 
substantially higher than Plan-and-Write (0.2374) and T-
CVAE (0.2606). It is a significant enhancement in lexical 
similarity and global narrative quality. The proposed system 
offers several key advantages: 
• Emotionally grounded narrative through emotion filtering 

with DistilRoBERTa. 
• Reduced redundancy and enhanced coherence by Fourier 

Transform-based FNet encoding. 
• There is better computational efficiency with reduced time 

complexity (O(n log n)) compared to the original 
transformer models (O(n²)).  

The following strengths render the system very applicable to 
real-world applications like AI-aided creative writing, 
individualized storytelling, frameworks for mental health 
support, and conversational agents with emotional 
intelligence. To further improve on these encouraging results, 
subsequent research will investigate the incorporation of 
more sophisticated emotion representation models and 
multi-modal inputs, such as visual or audio inputs. Further, 
more sophisticated evaluation metrics and human-in-the-
loop validation protocols will be employed to improve the 
narrative quality and emotional realism of the produced 
narratives. 

Ethical issue 
The authors are aware of and comply with best practices in 
publication ethics, specifically about authorship (avoidance of 
guest authorship), dual submission, manipulation of figures, 
competing interests, and compliance with policies on 
research ethics. The authors adhere to publication 
requirements and state that the submitted work is original 
and has not been published elsewhere. 

Data availability statement 
The manuscript contains all the data. However, more data will 
be available upon request from the authors. 

Conflict of interest 
The authors declare no potential conflict of interest. 

 
 



AH. Kachare et al. /Future Technology                                                                           November 2025| Volume 04 | Issue 04 | Pages 138-145 

144 
 

References 
[1] Jurafsky, D. (2000). Speech & language processing. 

Pearson Education India.  
ISBN-13: 9780131873216 

[2] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, 
L., Gomez, A. N., … Polosukhin, I. (2017). Attention is all 
you need. Advances in Neural Information Processing 
Systems, 30. 
ISBN: 9781510860964 

[3] Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., & 
Others. (2018). Improving language understanding by 
generative pre-training. 
Link: Click Here 

[4] Dwivedi, Y. K., Kshetri, N., Hughes, L., Slade, E. L., Jeyaraj, 
A., Kar, A. K., … Others. (2023). Opinion Paper: "So what 
if ChatGPT wrote it?" Multidisciplinary perspectives on 
opportunities, challenges and implications of 
generative conversational AI for research, practice and 
policy. International Journal of Information 
Management, 71, 102642. 
DOI: https://doi.org/10.1016/j.ijinfomgt.2023.102642 

[5] Brown, P. F. (1990). Class-based n-gram models of 
natural language. Comput. Linguist., 18, 18. 
Link: https://aclanthology.org/J92-4003/ 

[6] Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., 
Dhariwal, P., … Others. (2020). Language models are 
few-shot learners. Advances in Neural Information 
Processing Systems, 33, 1877–1901. 
ISBN: 9781713829546 

[7] Zhang, H., Song, H., Li, S., Zhou, M., & Song, D. (2023). A 
survey of controllable text generation using 
transformer-based pre-trained language models. ACM 
Computing Surveys, 56, 1–37. 
DOI: https://doi.org/10.1145/3617680 

[8] Meehan, J. R. (1976). The metanovel: writing stories by 
computer. Yale University. 
ISBN: 0824044096 

[9] Turner, S. R. (2014). The creative process: A computer 
model of storytelling and creativity. Psychology Press. 
DOI: https://doi.org/10.4324/9781315806464 

[10] Bringsjord, S., & Ferrucci, D. (1999). Artificial 
intelligence and literary creativity: Inside the mind of 
brutus, a storytelling machine. Psychology Press. 
DOI: https://doi.org/10.4324/9781410602398 

[11] Sutskever, I., Vinyals, O., & Le, Q. V. (2014). Sequence to 
sequence learning with neural networks. Advances in 
Neural Information Processing Systems, 27. 
ISBN: 9781510800410 

[12] Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., 
Sutskever, I., & Others. (2019). Language models are 
unsupervised multitask learners. OpenAI Blog, 1, 9. 

[13] Sepúlveda-Torres, R., Bonet-Jover, A., & Saquete, E. 
(2023). Detecting Misleading Headlines Through the 
Automatic Recognition of Contradiction in Spanish. 
IEEE Access, 11, 72007–72026. 
DOI: https://doi.org/10.1109/ACCESS.2023.3295781 

[14] Lebowitz, M. (1985). Story-telling as planning and 
learning. Poetics, 14, 483–502. 
DOI: https://doi.org/10.1016/0304-422X(85)90015-4 

[15] PÉrez, R. P. Ý., & Sharples, M. (2001). MEXICA: A 
computer model of a cognitive account of creative 

writing. Journal of Experimental & Theoretical Artificial 
Intelligence, 13, 119–139. 
DOI: https://doi.org/10.1080/09528130010029820 

[16] Riedl, M. O., & Young, R. M. (2010). Narrative planning: 
Balancing plot and character. Journal of Artificial 
Intelligence Research, 39, 217–268. 
DOI: https://doi.org/10.1613/jair.2989 

[17] Cavazza, M., Charles, F., & Mead, S. J. (2002). Character-
based interactive storytelling. IEEE Intelligent Systems, 
17, 17–24. 
DOI: https://doi.org/10.1109/MIS.2002.1024747 

[18] Fan, A., Lewis, M., & Dauphin, Y. (2018). Hierarchical 
Neural Story Generation. Proceedings of the 56th 
Annual Meeting of the Association for Computational 
Linguistics (Volume 1: Long Papers). 
DOI: https://doi.org/10.48550/arXiv.1805.04833 

[19] Xu, J., Ren, X., Zhang, Y., Zeng, Q., Cai, X., & Sun, X. (2018). 
A skeleton-based model for promoting coherence 
among sentences in narrative story generation. arXiv 
Preprint arXiv:1808. 06945. 
DOI: https://doi.org/10.48550/arXiv.1808.06945 

[20] Yao, L., Peng, N., Weischedel, R., Knight, K., Zhao, D., & 
Yan, R. (2019). Plan-and-write: Towards better 
automatic storytelling. Proceedings of the AAAI 
Conference on Artificial Intelligence, 33, 7378–7385. 
DOI: https://doi.org/10.1609/aaai.v33i01.33017378 

[21] Wang, T., & Wan, X. (2019). T-CVAE: Transformer-
based conditioned variational autoencoder for story 
completion. IJCAI, 5233–5239. 
DOI: https://doi.org/10.24963/ijcai.2019/727 

[22] Chen, G., Liu, Y., Luan, H., Zhang, M., Liu, Q., & Sun, M. 
(2020). Learning to generate explainable plots for 
neural story generation. IEEE/ACM Transactions on 
Audio, Speech, and Language Processing, 29, 585–593. 
DOI: https://doi.org/10.1109/TASLP.2020.3039606 

[23] Zhang, Y., Shi, X., Mi, S., & Yang, X. (2021). Image 
captioning with transformer and knowledge graph. 
Pattern Recognition Letters, 143, 43-49. 
DOI: https://doi.org/10.1016/j.patrec.2020.12.020 

[24] Chen, G., Liu, Y., Luan, H., Zhang, M., Liu, Q., & Sun, M. 
(2021). Learning to generate explainable plots for 
neural story generation. ACM Transactions on Audio, 
Speech, and Language Processing, 29, 585–593. 
DOI: https://doi.org/10.1109/TASLP.2020.3039606 

[25] Brahman, F., & Chaturvedi, S. (2020). Modeling 
protagonist emotions for emotion-aware storytelling. 
arXiv Preprint arXiv:2010. 06822. 
DOI: https://doi.org/10.48550/arXiv.2010.06822 

[26] Tan, B., Yang, Z., AI-Shedivat, M., Xing, E. P., & Hu, Z. 
(2020). Progressive generation of long text with 
pretrained language models. arXiv Preprint arXiv:2006. 
15720. 
DOI: https://doi.org/10.48550/arXiv.2006.15720 

[27] Min, K., Dang, M., & Moon, H. (2021). Deep learning-
based short story generation for an image using the 
encoder-decoder structure. IEEE Access, 9, 113550–
113557. 
DOI: https://doi.org/10.1109/ACCESS.2021.3104276 

[28] Wu, C., Wang, J., Yuan, S., Wang, L., & Zhang, W. (2021). 
Generate classical Chinese poems with theme-style 
from images. Pattern Recognition Letters, 149, 75–82. 

https://www.mikecaptain.com/resources/pdf/GPT-1.pdf
https://doi.org/10.1016/j.ijinfomgt.2023.102642
https://aclanthology.org/J92-4003/
https://doi.org/10.1145/3617680
https://doi.org/10.4324/9781315806464
https://doi.org/10.4324/9781410602398
https://doi.org/10.1109/ACCESS.2023.3295781
https://doi.org/10.1016/0304-422X(85)90015-4
https://doi.org/10.1080/09528130010029820
https://doi.org/10.1613/jair.2989
https://doi.org/10.1109/MIS.2002.1024747
https://doi.org/10.48550/arXiv.1805.04833
https://doi.org/10.48550/arXiv.1808.06945
https://doi.org/10.1609/aaai.v33i01.33017378
https://doi.org/10.24963/ijcai.2019/727
https://doi.org/10.1109/TASLP.2020.3039606
https://doi.org/10.1016/j.patrec.2020.12.020
https://doi.org/10.1109/TASLP.2020.3039606
https://doi.org/10.48550/arXiv.2010.06822
https://doi.org/10.48550/arXiv.2006.15720
https://doi.org/10.1109/ACCESS.2021.3104276


AH. Kachare et al. /Future Technology                                                                           November 2025| Volume 04 | Issue 04 | Pages 138-145 

145 
 

DOI: https://doi.org/10.1016/j.patrec.2021.05.016 
[29] Liu, Y., Huang, Q., Li, J., Mo, L., Cai, Y., & Li, Q. (2022). 

SSAP: Storylines and sentiment aware pre-trained 
model for story ending generation. IEEE/ACM 
Transactions on Audio, Speech, and Language 
Processing, 30, 686–694. 
DOI: https://doi.org/10.1109/TASLP.2022.3145320 

[30] Jin, Y., Kadam, V., & Wanvarie, D. (2022). Plot writing 
from pre-trained language models. arXiv Preprint 
arXiv:2206. 03021. 
DOI: https://doi.org/10.48550/arXiv.2206.03021 

[31] Chen, Y., Li, R., Shi, B., Liu, P., & Si, M. (2023). Visual story 
generation based on emotion and keywords. arXiv 
Preprint arXiv:2301. 02777. 
DOI: https://doi.org/10.48550/arXiv.2301.02777 

[32] Khan, L. P., Gupta, V., Bedi, S., & Singhal, A. (2023). 
StoryGenAI: An Automatic Genre-Keyword Based Story 
Generation. 2023 International Conference on 
Computational Intelligence and Sustainable 
Engineering Solutions (CISES), 955–960. 
DOI: 
https://doi.org/10.1109/CISES58720.2023.10183482 

[33] Hartmann, J. (2022). Emotion english distilroberta-
base. See  
Link: https://huggingface.co/j-hartmann/emotion-
english-distilroberta-base 

[34] Lee-Thorp, J., Ainslie, J., Eckstein, I., & Ontanon, S. 
(2021). Fnet: Mixing tokens with fourier transforms. 
arXiv preprint arXiv:2105.03824. 
DOI: https://doi.org/10.48550/arXiv.2105.03824 

[35] Fu, K., Li, H., & Shi, X. (2024). An encoder-decoder 
architecture with Fourier attention for chaotic time 
series multi-step prediction. Applied Soft Computing, 
156(111409), 111409. 
DOI: https://doi.org/10.1016/j.asoc.2024.111409 

[36] Dittakan, K., Prompitak, K., Thungklang, P., & 
Wongwattanakit, C. (2023). Image caption generation 
using transformer learning methods: a case study on 
instagram image. Multimedia Tools and Applications, 
83(15), 46397–46417. 
DOI: https://doi.org/10.1007/s11042-023-17275-9 

[37] Danescu-Niculescu-Mizil, C., & Lee, L. (2011). 
Chameleons in imagined conversations: A new 
approach to understanding coordination of linguistic 
style in dialogs. arXiv Preprint arXiv:1106. 3077. 
DOI: https://doi.org/10.48550/arXiv.1106.3077 

[38] Zhu, Y. (2024). Visual7W dataset [Data set]. 
DOI: https://doi.org/10.57702/zqariweh 

[39] Mostafazadeh, N. (2024). ROCStories [Data set]. 
DOI: https://doi.org/10.57702/26yy027v 

[40] Lee, S., Lee, J., Moon, H., Park, C., Seo, J., Eo, S., … Lim, H. 
(2023). A survey on evaluation metrics for machine 
translation. Mathematics, 11(4), 1006. 
DOI: https://doi.org/10.3390/math11041006 

[41] Kaptein, F., & Broekens, J. (2015, August). The affective 
storyteller: using character emotion to influence 
narrative generation. In International Conference on 
Intelligent Virtual Agents (pp. 352-355). Cham: 
Springer International Publishing. 
DOI: http://doi.org/10.1007/978-3-319-21996-7_38 

[42] Rashkin, H., Celikyilmaz, A., Choi, Y., & Gao, J. (2020). 
PlotMachines: Outline-conditioned generation with 
dynamic plot state tracking. arXiv preprint 
arXiv:2004.14967. 
DOI: https://doi.org/10.48550/arXiv.2004.14967 

[43] Li, Y., Gan, Z., Shen, Y., Liu, J., Cheng, Y., Wu, Y., ... & Gao, 
J. (2019). Storygan: A sequential conditional gan for 
story visualization. In Proceedings of the IEEE/CVF 
conference on computer vision and pattern recognition 
(pp. 6329-6338). 
DOI: https://doi.org/10.48550/arXiv.1812.02784 
 

  

This article is an open-access article distributed under the 
terms and conditions of the Creative Commons Attribution 
(CC BY) license 
(https://creativecommons.org/licenses/by/4.0/). 

https://doi.org/10.1016/j.patrec.2021.05.016
https://doi.org/10.1109/TASLP.2022.3145320
https://doi.org/10.48550/arXiv.2206.03021
https://doi.org/10.48550/arXiv.2301.02777
https://doi.org/10.1109/CISES58720.2023.10183482
https://huggingface.co/j-hartmann/emotion-english-distilroberta-base
https://huggingface.co/j-hartmann/emotion-english-distilroberta-base
https://doi.org/10.48550/arXiv.2105.03824
https://doi.org/10.1016/j.asoc.2024.111409
https://doi.org/10.1007/s11042-023-17275-9
https://doi.org/10.48550/arXiv.1106.3077
https://doi.org/10.57702/zqariweh
https://doi.org/10.57702/26yy027v
https://doi.org/10.3390/math11041006
http://doi.org/10.1007/978-3-319-21996-7_38
https://doi.org/10.48550/arXiv.2004.14967
https://doi.org/10.48550/arXiv.1812.02784
https://creativecommons.org/licenses/by/4.0/

	1. Introduction
	Text generation is a process of automatically creating text that sounds like it was written by a human, which is trending in the world of artificial intelligence (AI). It has its roots in the early days of computational linguistics research [1]. Nowad...
	2. Background
	The historical trajectory of story generation is deeply rooted in human culture, evolving from oral traditions and written literature to computational approaches [8]. Early computer programs utilized rule-based systems, followed by more sophisticated ...
	3. Literature survey
	The research conducted by Fan A. and Lewis M. [18] aimed to enhance fluency and coherence by utilizing a hierarchical model with self-attention, but encountered issues such as tokenization and text repetition. Xu J. and Ren X. [19] focused on making s...
	 Failure to manage the proper emotional context: Most models fail to guarantee that the output text is always tied to a target emotion throughout the story.
	 Inadequate modelling of long-range dependencies: Current systems do not have strong mechanisms for modelling and preserving coherence across sentences or pieces of a story.
	 Recurrent behaviours and narrative flow inconsistencies are commonly found in generative models because of shallow attention mechanisms or the absence of semantic feedback loops.
	 Lack of a cohesive system that identifies emotions and increases coherence: Few research efforts have attempted to combine emotional filtering with semantic structuring in a single, effective system.
	These identified gaps highlight the need for a unified story generation system to simultaneously manage emotional control and narrative coherence using efficient, scalable techniques. Although deep learning models like GPT-2 have revolutionized natura...
	 To ensure emotional alignment by filtering GPT-2-generated sentences using DistilRoBERTa.
	 To enhance narrative coherence by modeling global dependencies through a Fourier-based FNET architecture.
	 To empirically validate the system on benchmark datasets (Cornell Movie Dialogs, Visual7W, and ROCStories).
	 To evaluate performance using BLEU and METEOR metrics and emotion retention accuracy.
	4. Proposed system
	Having performed an extensive review of existing literature, we identify two long-standing issues in the domain of automatic narrative generation:
	 Repetition and Inconsistency: The majority of modern systems lack consistent character behavior and consistent story progression, instead generating repetitive or contradictory content.
	 Lack of Cross-Sentence Dependency Modeling: The current models are limited by their failure to capture long-distance semantic and affective relations between sentences, impacting narrative coherence.
	We have developed a hybrid model, as shown in Figure 1, that combines GPT-2, DistilRoBERTa, and a specially crafted FNET model to tackle specific challenges. FNET is a Transformer encoder that mixes via the Fourier transforms instead of using the usua...
	Figure 1. Proposed system for short story generation
	The reason to use FNET:
	 Global Context Awareness: Many attention mechanisms become complex as the input grows. Nevertheless, FNET's Fourier transform encoder is more scalable. It can handle complete sequences.
	 Improved Consistency: Another great thing about FNET is that it keeps the narrative on track. It minimizes those annoying moments where the story drifts or characters seem inconsistent. It does this by holding onto those long-range dependencies.
	 Reduced Computational Expense: FNET skips the whole attention mechanism altogether. It operates with a time complexity of O(n log n), which is significantly better than the O(n²) we see in standard Transformers. The system, as proposed, works in two...
	A comparative analysis of FNet with Linformer, Longformer, and Performer is presented in Table 1, which outlines the trade-offs across computational efficiency, memory footprint, and real-world performance. It supports the design decision to employ FN...
	4.1 Stage 1: Emotionally aware sentence generation
	 The system is presented with a question, a corresponding answer, and a target emotional state (e.g., happiness, fear, anger).
	 Sentence Generation: A pre-trained GPT-2 model produces some potential sentences based on the inputs provided.
	 Emotion Filtering: Every sentence is passed through a DistilRoBERTa-based emotion classifier. Sentences belonging to the target emotion only are kept.
	 Selection Mechanism: A beam search algorithm produces diverse sentence forms with maximum semantic and emotional appropriateness.
	We employed the Emotion English DistilRoBERTa‑base model [33], classifying text into seven categories: anger, disgust, fear, joy, neutral, sadness, and surprise. It is built by fine‑tuning DistilRoBERTa‑base on a balanced subset (~20k samples) drawn f...
	4.2 Stage 2: FNET-based coherent story generation
	The FNet encoder replaces the standard self-attention mechanism of transformers with a 2D Discrete Fourier Transform (DFT2), achieving efficient global mixing with reduced computational complexity. This section formally describes the mathematical oper...
	𝑋 ∈ ,𝑅-𝑛×𝑑.                              (1)
	Where n: sequence length, d: embedding dimension, each row ,𝑥-𝑖. correspond to the embedding of the ith token.
	To preserve the order of tokens in the input sequence, we add fixed sinusoidal positional encodings 𝑃 ∈ ,𝑅-𝑛×𝑑.and type encodings 𝑇 ∈ ,𝑅-𝑛×𝑑.
	,𝑋-′.=𝑋+𝑃+𝑇                                                            (2)
	These encodings allow the model to distinguish between tokens at different positions without learning position-specific parameters. Instead of using quadratic-complexity self-attention, FNet applies a 2D Discrete Fourier Transform (DFT2) across both t...
	𝑍=𝑅𝑒 ,𝐹𝐹𝑇2,,𝑋-′...,𝑅-𝑛×𝑑.                                              (3)
	Table 1. Comparative analysis of efficient transformer variants across time complexity, memory usage, parallelizability, and practical applicability
	Here, FFT2(.) is a 2D Fourier Transform applied to the matrix (1st DFT across Rows: Sequence dimension and then 2nd DFT across columns: Feature dimension), and Re(.) extracts the real part to keep the result compatible with downstream layers. This ope...
	𝑌=𝐿𝑎𝑦𝑒𝑟𝑁𝑜𝑟𝑚 ,𝑍+,𝑋-′..                                              (4)
	The normalized output is passed through a standard position-wise feedforward network:
	𝐻=𝑅𝑒𝐿𝑈,,𝑌𝑊-1.+,𝑏-1..,𝑊-2.+,𝑏-2.                                                   (5)
	Where, ,𝑊-1.,,𝑊-2. ∈ ,𝑅-𝑛×𝑑. are learnable weight matrices and ,𝑏-1., ,𝑏-2. ∈ ,𝑅-𝑑. are biases. ReLU adds non-linearity to improve representational power.
	Finally, another residual connection is added between the input to the Feedforward block and the output of the Feedforward block. It helps the model combine local (feedforward) and global (Fourier) features. After adding them, LayerNorm is applied aga...
	,𝐻-′.=𝐿𝑎𝑦𝑒𝑟𝑁𝑜𝑟𝑚 ,𝑌+𝐻.                                    (6)
	Once the final representation ,𝐻-′. is ready and mapped to the vocabulary space using a linear output projection. It is done by multiplying it with a learnable weight matrix:
	𝐿𝑜𝑔𝑖𝑡𝑠=,𝐻-′.,𝑊-0.+,𝑏-0.                                     (7)
	The model applies the softmax function to the logits to generate the next token and get probabilities. Then, it uses argmax to choose the most likely token:
	,,𝑦.-𝑡.=,arg-,max-,𝑆𝑜𝑓𝑡𝑚𝑎𝑥,𝐿𝑜𝑔𝑖𝑡𝑠....                         (8)
	The final output is a short story composed of emotionally consistent and semantically connected sentences generated from the filtered GPT-2 outputs and refined through FNET's encoding-decoding mechanism.
	This system leverages FNET to pick up on the subtle meanings and feelings, along with GPT-2's amazing storytelling abilities. Such a multi-step journey; craft short stories that resonate with emotion while responding to specific questions and answers....
	Figure 2. FNET model
	5. Evaluation of the proposed system
	We evaluate the system to see how well it could write an emotionally engaging short story. We examine three different benchmark datasets, each one having some unique challenges. The system was evaluated against some established baseline models to meas...
	5.1 Dataset
	So, we decided to work with three benchmark datasets for our experiments to see how well the proposed system performs. First up, we have the Cornell Movie-Dialogs Corpus [37]. It is a pretty massive collection, with all these fictional conversations p...
	5.2 Baseline models
	We compare our models with the following: 1) T-CVAE [21] leverages transformers and a variational autoencoder to learn story patterns and generate new, coherent stories, with the ability to incorporate additional input for guidance. 2) Plan&Write [20]...
	5.3 Metrics
	We have chosen to utilize the BLEU (Bilingual Evaluation Understudy) [40] and METEOR [40] metrics to evaluate the quality of the generated stories. BLEU is one of the most popular machine translation metrics that determines the similarity between a hu...
	The BLEU score formula is derived based on precision and a penalty for brevity. The formula is:
	𝐵𝐿𝐸𝑈=𝐵𝑃 ×,exp-,,1-𝑛.,𝑖=1-𝑛-,log-,𝑝-𝑖.....                                 (9)
	However, BLEU does not consider the stems and synonyms of words. To overcome these limitations, METEOR uses a weighted F1-score and penalty function.
	,𝐹-𝑆𝐶𝑂𝑅𝐸.=,10,,𝑚𝑎𝑡𝑐ℎ𝑒𝑑 𝑢𝑛𝑖𝑔𝑟𝑎𝑚-𝑢𝑛𝑖𝑔𝑟𝑎𝑚 𝑖𝑛 ℎ𝑦𝑝𝑜𝑡ℎ𝑒𝑖𝑠..,,𝑚𝑎𝑡𝑐ℎ𝑒𝑑 𝑢𝑛𝑖𝑔𝑟𝑎𝑚-𝑢𝑛𝑖𝑔𝑟𝑎𝑚 𝑖𝑛𝑟𝑒𝑓𝑒𝑟𝑒𝑛𝑐𝑒..-,,𝑚𝑎𝑡𝑐ℎ𝑒𝑑 𝑢𝑛𝑖𝑔𝑟𝑎𝑚-𝑢𝑛𝑖𝑔𝑟𝑎𝑚 𝑖𝑛𝑟𝑒𝑓𝑒𝑟𝑒𝑛𝑐𝑒..+9,,𝑚𝑎𝑡𝑐ℎ𝑒𝑑 𝑢𝑛...
	𝑃𝑒𝑛𝑎𝑙𝑡𝑦=0.5 × ,,𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐ℎ𝑢𝑛𝑘𝑠 𝑖𝑛 𝐻𝑦𝑝𝑜𝑡ℎ𝑒𝑖𝑠-𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑚𝑎𝑡𝑐ℎ𝑒𝑑 𝑢𝑛𝑖𝑔𝑟𝑎𝑚𝑠..                 (11)
	𝑀𝐸𝑇𝐸𝑂𝑅=,𝐹-𝑆𝐶𝑂𝑅𝐸. × ,1−𝑃𝑒𝑛𝑎𝑙𝑡𝑦.                   (12)
	5.4 Experimental setup
	We utilize a pre-trained GPT-2 model to generate a wide range of sentences. It will take a question and its answer, then churn out several ways to say it. Now, the emotion classifier will look at all those candidate sentences and determine which hits ...
	5.5 Result
	This section presents the experimental evaluation's outcomes, showcasing the proposed system's performance in generating emotionally aware short stories. The results are analyzed in comparison to baseline models, highlighting the strengths and limitat...
	 Cornell's Movie Dataset: This dataset likely contains movie dialogues or scripts. The FNET model achieved an accuracy of 0.60 (60%) on this dataset. It suggests that the model performs moderately well in understanding or generating movie dialogue. T...
	 Visual7w Dataset: This dataset probably contains images paired with questions and answers about the visual content of the images. The FNET model achieved a high accuracy of 0.85 (85%) on this dataset, indicating that it is quite effective at underst...
	 ROC-Stories: This dataset likely contains short stories or narratives. FNET achieved an accuracy of 0.67 (67%) on this dataset. It suggests that the model performs reasonably well in understanding or generating short stories, although there is still...
	Table 2. Accuracy of trained FNet on different benchmark datasets
	5.6 Discussion
	Table 3 compares three models, Plan-and-Write, T-CVAE, and the Proposed System, based on BLEU and METEOR metrics, which evaluate how closely the generated text matches the reference human-written text.
	BLEU score analysis
	 Plan-and-Write: This model has the lowest scores across all BLEU metrics. It suggests that its generated text has the least overlap in individual words, bigrams, and trigrams compared to the human-generated reference text.
	 T-Cave: It shines regarding BLEU-2 scores, which means it is pretty good at churning out pairs of words matching the reference text. However, its scores for BLEU-1. It hints that it might struggle with individual words.
	 The Proposed model: It grabs the top spot for BLEU-1, which means it nails those individual word matches well. Its BLEU-2 score is not the absolute highest, but they are still solid compared to other models.
	 Overall: So, T-Cave is excellent for generating those word pairs, but the Proposed model seems to strike a better balance across all the BLEU metrics, which suggests it might come closer to sounding like a human would write, especially regarding voc...
	METEOR score analysis
	 Plan-and-Write achieves a METEOR score of 0.266, indicating the lowest semantic alignment among the three.
	 T-CVAE slightly improves with a METEOR of 0.278, reflecting a modest gain in semantic closeness.
	 The Proposed System obtains the highest METEOR score of 0.291, indicating better overall alignment with human-written references, not just in exact word matches, but also in meaning and structure.
	Table 3. Comparative analysis of BLUE and METEOR scores for baseline and proposed system for ROCStories dataset
	Since the output generations from T-CVAE and Plan-and-Write were not publicly available, we were unable to perform statistical significance testing. However, our proposed model consistently outperforms the baselines across multiple metrics and dataset...
	Comparison with other existing emotional story generation models:
	FNet-GPT differentiates itself from AffectStory [41], PlotMachines [42], and StoryGAN [43] by providing a unique blend of explicit emotional control, enhanced textual coherence, and superior computational efficiency for narrative generation. While Aff...
	5.7 Error analysis
	Despite employing an emotion classifier (DistilRoBERTa) during the filtering stage, we found that a few generated stories from the Cornell dataset deviated from the intended emotional tone. Similarly, in all datasets, a few outputs showed weak narrati...
	6. Conclusion
	This paper introduces a novel, emotionally intelligent short story generation architecture by combining the robust language generation ability of GPT-2 with the semantic richness and sparsity of a specially designed FNet model. The combined model addr...
	 Emotionally grounded narrative through emotion filtering with DistilRoBERTa.
	 Reduced redundancy and enhanced coherence by Fourier Transform-based FNet encoding.
	 There is better computational efficiency with reduced time complexity (O(n log n)) compared to the original transformer models (O(n²)).
	The following strengths render the system very applicable to real-world applications like AI-aided creative writing, individualized storytelling, frameworks for mental health support, and conversational agents with emotional intelligence. To further i...
	The manuscript contains all the data. However, more data will be available upon request from the authors.
	Conflict of interest
	The authors declare no potential conflict of interest.
	References
	[1] Jurafsky, D. (2000). Speech & language processing. Pearson Education India.
	ISBN-13: 9780131873216
	[2] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.
	ISBN: 9781510860964
	[3] Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., & Others. (2018). Improving language understanding by generative pre-training.
	Link: Click Here
	[4] Dwivedi, Y. K., Kshetri, N., Hughes, L., Slade, E. L., Jeyaraj, A., Kar, A. K., … Others. (2023). Opinion Paper: "So what if ChatGPT wrote it?" Multidisciplinary perspectives on opportunities, challenges and implications of generative conversation...
	DOI: https://doi.org/10.1016/j.ijinfomgt.2023.102642
	[5] Brown, P. F. (1990). Class-based n-gram models of natural language. Comput. Linguist., 18, 18.
	Link: https://aclanthology.org/J92-4003/
	[6] Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., … Others. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33, 1877–1901.
	ISBN: 9781713829546
	[7] Zhang, H., Song, H., Li, S., Zhou, M., & Song, D. (2023). A survey of controllable text generation using transformer-based pre-trained language models. ACM Computing Surveys, 56, 1–37.
	DOI: https://doi.org/10.1145/3617680
	[8] Meehan, J. R. (1976). The metanovel: writing stories by computer. Yale University.
	ISBN: 0824044096
	[9] Turner, S. R. (2014). The creative process: A computer model of storytelling and creativity. Psychology Press.
	DOI: https://doi.org/10.4324/9781315806464
	[10] Bringsjord, S., & Ferrucci, D. (1999). Artificial intelligence and literary creativity: Inside the mind of brutus, a storytelling machine. Psychology Press.
	DOI: https://doi.org/10.4324/9781410602398
	[11] Sutskever, I., Vinyals, O., & Le, Q. V. (2014). Sequence to sequence learning with neural networks. Advances in Neural Information Processing Systems, 27.
	ISBN: 9781510800410
	[12] Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., & Others. (2019). Language models are unsupervised multitask learners. OpenAI Blog, 1, 9.
	[13] Sepúlveda-Torres, R., Bonet-Jover, A., & Saquete, E. (2023). Detecting Misleading Headlines Through the Automatic Recognition of Contradiction in Spanish. IEEE Access, 11, 72007–72026.
	DOI: https://doi.org/10.1109/ACCESS.2023.3295781
	[14] Lebowitz, M. (1985). Story-telling as planning and learning. Poetics, 14, 483–502.
	DOI: https://doi.org/10.1016/0304-422X(85)90015-4
	[15] PÉrez, R. P. Ý., & Sharples, M. (2001). MEXICA: A computer model of a cognitive account of creative writing. Journal of Experimental & Theoretical Artificial Intelligence, 13, 119–139.
	DOI: https://doi.org/10.1080/09528130010029820
	[16] Riedl, M. O., & Young, R. M. (2010). Narrative planning: Balancing plot and character. Journal of Artificial Intelligence Research, 39, 217–268.
	DOI: https://doi.org/10.1613/jair.2989
	[17] Cavazza, M., Charles, F., & Mead, S. J. (2002). Character-based interactive storytelling. IEEE Intelligent Systems, 17, 17–24.
	DOI: https://doi.org/10.1109/MIS.2002.1024747
	[18] Fan, A., Lewis, M., & Dauphin, Y. (2018). Hierarchical Neural Story Generation. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).
	DOI: https://doi.org/10.48550/arXiv.1805.04833
	[19] Xu, J., Ren, X., Zhang, Y., Zeng, Q., Cai, X., & Sun, X. (2018). A skeleton-based model for promoting coherence among sentences in narrative story generation. arXiv Preprint arXiv:1808. 06945.
	DOI: https://doi.org/10.48550/arXiv.1808.06945
	[20] Yao, L., Peng, N., Weischedel, R., Knight, K., Zhao, D., & Yan, R. (2019). Plan-and-write: Towards better automatic storytelling. Proceedings of the AAAI Conference on Artificial Intelligence, 33, 7378–7385.
	DOI: https://doi.org/10.1609/aaai.v33i01.33017378
	[21] Wang, T., & Wan, X. (2019). T-CVAE: Transformer-based conditioned variational autoencoder for story completion. IJCAI, 5233–5239.
	DOI: https://doi.org/10.24963/ijcai.2019/727
	[22] Chen, G., Liu, Y., Luan, H., Zhang, M., Liu, Q., & Sun, M. (2020). Learning to generate explainable plots for neural story generation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29, 585–593.
	DOI: https://doi.org/10.1109/TASLP.2020.3039606
	[23] Zhang, Y., Shi, X., Mi, S., & Yang, X. (2021). Image captioning with transformer and knowledge graph. Pattern Recognition Letters, 143, 43-49.
	DOI: https://doi.org/10.1016/j.patrec.2020.12.020
	[24] Chen, G., Liu, Y., Luan, H., Zhang, M., Liu, Q., & Sun, M. (2021). Learning to generate explainable plots for neural story generation. ACM Transactions on Audio, Speech, and Language Processing, 29, 585–593.
	DOI: https://doi.org/10.1109/TASLP.2020.3039606
	[25] Brahman, F., & Chaturvedi, S. (2020). Modeling protagonist emotions for emotion-aware storytelling. arXiv Preprint arXiv:2010. 06822.
	DOI: https://doi.org/10.48550/arXiv.2010.06822
	[26] Tan, B., Yang, Z., AI-Shedivat, M., Xing, E. P., & Hu, Z. (2020). Progressive generation of long text with pretrained language models. arXiv Preprint arXiv:2006. 15720.
	DOI: https://doi.org/10.48550/arXiv.2006.15720
	[27] Min, K., Dang, M., & Moon, H. (2021). Deep learning-based short story generation for an image using the encoder-decoder structure. IEEE Access, 9, 113550–113557.
	DOI: https://doi.org/10.1109/ACCESS.2021.3104276
	[28] Wu, C., Wang, J., Yuan, S., Wang, L., & Zhang, W. (2021). Generate classical Chinese poems with theme-style from images. Pattern Recognition Letters, 149, 75–82.
	DOI: https://doi.org/10.1016/j.patrec.2021.05.016
	[29] Liu, Y., Huang, Q., Li, J., Mo, L., Cai, Y., & Li, Q. (2022). SSAP: Storylines and sentiment aware pre-trained model for story ending generation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30, 686–694.
	DOI: https://doi.org/10.1109/TASLP.2022.3145320
	[30] Jin, Y., Kadam, V., & Wanvarie, D. (2022). Plot writing from pre-trained language models. arXiv Preprint arXiv:2206. 03021.
	DOI: https://doi.org/10.48550/arXiv.2206.03021
	[31] Chen, Y., Li, R., Shi, B., Liu, P., & Si, M. (2023). Visual story generation based on emotion and keywords. arXiv Preprint arXiv:2301. 02777.
	DOI: https://doi.org/10.48550/arXiv.2301.02777
	[32] Khan, L. P., Gupta, V., Bedi, S., & Singhal, A. (2023). StoryGenAI: An Automatic Genre-Keyword Based Story Generation. 2023 International Conference on Computational Intelligence and Sustainable Engineering Solutions (CISES), 955–960.
	DOI: https://doi.org/10.1109/CISES58720.2023.10183482
	[33] Hartmann, J. (2022). Emotion english distilroberta-base. See
	Link: https://huggingface.co/j-hartmann/emotion-english-distilroberta-base
	[34] Lee-Thorp, J., Ainslie, J., Eckstein, I., & Ontanon, S. (2021). Fnet: Mixing tokens with fourier transforms. arXiv preprint arXiv:2105.03824.
	DOI: https://doi.org/10.48550/arXiv.2105.03824
	[35] Fu, K., Li, H., & Shi, X. (2024). An encoder-decoder architecture with Fourier attention for chaotic time series multi-step prediction. Applied Soft Computing, 156(111409), 111409.
	DOI: https://doi.org/10.1016/j.asoc.2024.111409
	[36] Dittakan, K., Prompitak, K., Thungklang, P., & Wongwattanakit, C. (2023). Image caption generation using transformer learning methods: a case study on instagram image. Multimedia Tools and Applications, 83(15), 46397–46417.
	DOI: https://doi.org/10.1007/s11042-023-17275-9
	[37]  Danescu-Niculescu-Mizil, C., & Lee, L. (2011). Chameleons in imagined conversations: A new approach to understanding coordination of linguistic style in dialogs. arXiv Preprint arXiv:1106. 3077.
	DOI: https://doi.org/10.48550/arXiv.1106.3077
	[38] Zhu, Y. (2024). Visual7W dataset [Data set].
	DOI: https://doi.org/10.57702/zqariweh
	[39] Mostafazadeh, N. (2024). ROCStories [Data set].
	DOI: https://doi.org/10.57702/26yy027v
	[40] Lee, S., Lee, J., Moon, H., Park, C., Seo, J., Eo, S., … Lim, H. (2023). A survey on evaluation metrics for machine translation. Mathematics, 11(4), 1006.
	DOI: https://doi.org/10.3390/math11041006
	[41] Kaptein, F., & Broekens, J. (2015, August). The affective storyteller: using character emotion to influence narrative generation. In International Conference on Intelligent Virtual Agents (pp. 352-355). Cham: Springer International Publishing.
	DOI: http://doi.org/10.1007/978-3-319-21996-7_38
	[42] Rashkin, H., Celikyilmaz, A., Choi, Y., & Gao, J. (2020). PlotMachines: Outline-conditioned generation with dynamic plot state tracking. arXiv preprint arXiv:2004.14967.
	DOI: https://doi.org/10.48550/arXiv.2004.14967
	[43] Li, Y., Gan, Z., Shen, Y., Liu, J., Cheng, Y., Wu, Y., ... & Gao, J. (2019). Storygan: A sequential conditional gan for story visualization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 6329-6338).
	DOI: https://doi.org/10.48550/arXiv.1812.02784

