Asadullah Yousaf
← Featured work
Research05 • Research • 2024

Story Generation: LSTM vs GRU

Comparative sequence modeling for outline-to-story generation.

Research pipeline comparing LSTM and GRU models for interactive, seed-word story generation on classic short stories.

Stack

6 technologies

Gallery

2 screenshots

Source

Open on GitHub

Problem

For outline-to-story generation, teams need evidence on which recurrent architecture balances coherence, creativity, and training efficiency before investing in production NLP pipelines.

Solution

Designed a comparative research pipeline that preprocesses classic short-story corpora, trains bidirectional LSTM and GRU models on shared outline-to-story tasks, and evaluates outputs on coherence, creativity, and convergence behavior.

Highlights

  • Controlled comparison between LSTM and GRU under identical datasets
  • Bidirectional LSTM with 256 hidden units and interactive seed-word prediction
  • Evaluation focused on quality, creativity, and training efficiency

Outcomes

  • 79% accuracy based on human evaluation (per project README)
  • Reproducible baseline for sequence-model selection
  • Documented trade-offs useful for future generative text systems

Repository

Story Generation repository

LSTM and GRU models for interactive story generation with configurable training and human-evaluated outputs.

Human evaluation

79% accuracy

LSTM architecture

Bidirectional, 256 hidden units

Dataset

Classic short stories (ClassicShorts)

Training scope

100,000 instances per README

  • LSTM and GRU implementations for story generation
  • Interactive prediction with user-provided seed words and context
  • Configurable sequence length and context vectors
  • Jupyter notebooks for training and evaluation workflows
  • Flask app for interactive story generation

Stack

PythonKerasLSTMGRUNumPyGensim