Story Generation using LSTM & GRU

Project Overview

Machines, powered by advancements in Natural Language Processing (NLP), can now autonomously weave stories with contextual understanding. This project leverages NLP tools and libraries to build a model capable of automatically generating rich stories. Text generation, a branch of NLP, finds applications in various domains such as automatic documentation systems, letter writing, and report generation.

Dataset

The project utilizes a dataset obtained from a website filled with stories. You can find the dataset here.

Methodology

• Data preprocessing involves removing punctuation and using sentences with 7 or more words for training.
• Training sets are limited to 100,000 instances to manage compute memory space.
• Input and output training instances are generated from the dataset.
• All words are vectorized using one-hot coding.
• The model uses a bidirectional LSTM network with 256 hidden units, "relu" activation in the inner layer, and "softmax" activation in the outer layer.
• Keras library is employed for training the model.
• Implementation is done in Python using libraries such as Scikit-learn, NumPy, and Gensim.

Model Configuration

• Input sequence length (excluding context vector): 8
• Context vector: V-dimensional vector (size of vocabulary)
• Configuration: All-to-one configuration, where the input contains a context vector, and the output contains the next word in the sequence.

Prediction Phase

• Two sets of input from the user: sequence of seed words and context word(s).
• The network assembles input and predicts subsequent words.
• The process continues until the sentence is completed, indicated by dots.
• The next context word is then used for generating the next sentence.

Evaluation

• The model minimizes drag loss by adjusting parameters like the number of neurons, layers, stack size, and sequence length.
• Human evaluation achieved 79% accuracy.
• Further improvements can be made by considering the meaning of words in context and incorporating synonyms.

Future Extensions

The system can be extended for automatic generation of news, news articles, jokes, or posts.

Dependencies

• Python
• Scikit-learn
• NumPy
• Gensim

How to Use

1. Install the required dependencies.
2. Download the dataset from here and place it in the appropriate directory.
3. Run the code to train the model.
4. Provide seed words and context words during the prediction phase.

Happy Story Generation!